The match type is dead. Long live the keyword.

I feel like I’m watching one funeral after another for keywords. Every post, hot take, and panel is eulogizing the same thing: the search keyword, dead at the hands of AI automation. I’m in the camp that thinks folks are mourning the wrong casualty.

What’s being buried is the match type, and specifically the belief that choosing exact over phrase ever gave you mechanical control over what triggers your ads. The keyword survives because the intent behind it, reaching the people searching for a specific thing, is still valid and still achievable. What’s ending is the assumption that the match type you pick enforces that intent with any kind of precision.

What Google actually said at GML 2026

Vidhya Srinivasan, VP of Google Ads, was direct about where this is heading: “You can’t choose keywords anymore.” Most people have read that as proof of a keywordless future. If you read it again, you’ll see that it’s proof of a match-typeless one, which is a very different claim with strategic implications to match.

The mechanical evidence is already sitting in your campaign settings. When you enable the broad match campaign setting, Google converts all your existing phrase and exact match keywords into broad and applies broad to everything you add afterward. That’s Google making a deliberate platform-level decision. And Dynamic Search Ads will automatically migrate to AI Max for Search starting in September 2026, which quietly removes the last query-discovery workaround a lot of us had been leaning on.

On the control side, Google is building a steering layer for its automated campaign types. The clearest example is AI Brief, announced for AI Max and Performance Max, which lets you guide the system in plain language across three layers: messaging (what your ads should and shouldn’t say), matching (which searches to chase or avoid), and audience (who to reach and how to speak to them). Matching still exists; it’s just an instruction you write out now, something like “prioritize searches for healthy pantry staples,” rather than a match type you assign to a keyword. The real job of all this is to make advertisers comfortable enough with automatically generated assets and final URL expansion to actually hand them over (or at least that’s my theory). You’re no longer programming a lookup table; you’re briefing an inference engine and reviewing its previews before anything goes live.

What that adds up to is a genuinely useful control surface that works nothing like a match type; the control now lives in how you brief and review the system rather than in the exact-versus-phrase decision you used to agonize over.

 

Why the keyword survives

 

The “keywords are dead” narrative has a surface-level logic to it, which is exactly why it spreads so easily.

Kirk Williams documented what’s actually happening to search behaviour, and it reads as a bifurcation (something dividing into two) rather than a collapse. Queries are splitting into two distinct patterns. Long, conversational, multi-part prompts are showing up inside AI Mode, where Google’s own one-year data puts the average query at three times the length of a traditional search. No keyword list can predict or capture that territory in advance. But the short, high-intent transactional queries, “emergency HVAC Chicago,” “three bedroom townhome Langley BC,” “Invisalign cost Los Angeles,” are still being typed, still resolving to immediate action, and still need something that functions like a keyword to anchor the match.

The AI handles exploration; the keyword handles the transaction. That’s how people use search when they’re ready to spend money or pick up the phone, and it holds regardless of how good the exploration layer gets. The algorithm can’t surface a relevant result for a local service query out of ambient signal alone. It needs a thematic seed, and the keyword is that seed.

Which is exactly why the whole game moves to signal quality. The keyword is one signal, the conversion you optimize toward is another, and the system will chase whatever you hand it, faithfully, whether or not it’s the thing you actually wanted. The cleanest example of that I’ve seen came down to a single misconfigured conversion. An insurance agent friend, Mary, asked me to look at her account because something felt off, even while the numbers looked incredible: conversions up 416% month over month, cost per lead down 21%, clicks up 23%, impressions down 32%. On a dashboard, that combination reads as a campaign that’s crushing it. What was actually happening was that her one primary conversion, the signal telling Google to go find more, fired the second someone started typing in the contact form rather than when they submitted it. So the system did exactly what it was told and went looking for people who would touch a keyboard, not people who would buy a mortgage. Reported conversions climbed while the real business moved the other way, and nothing in the account flagged the signal as wrong, because as far as the system was concerned, it wasn’t; it was optimizing perfectly toward the target it had been handed.

That’s automation drift: the system optimizing flawlessly toward a target that was quietly wrong, with the dashboard green the entire time. It’s the failure mode that gets more dangerous as match types flatten, because the wall that used to contain a wrong signal is coming down.

 

What signal governance actually looks like

 

If the matching logic now lives inside the model instead of in your match-type setup, your real control has moved upstream. Exact match used to act as a hard wall that physically stopped the system from wandering into irrelevant queries; inside AI Max, that wall is gone. Your exact and phrase keywords get treated as broad, and Mike Ryan’s data puts exact-match expansion at between 27% and 89% of AI Max match type impressions, averaging around 80%. Once that containment disappears, the system’s behaviour is downstream of one thing: the quality of the conversion signal you feed it. That’s where lead gen is most exposed, because its conversions are thin. A form fill or a phone call is a single event with no built-in proof that anything valuable happened, so when the system goes looking for more of them, it takes the path of least resistance, optimizing toward the cheapest version of that event rather than the customer behind it. Andrew Lolk has made the related case that e-commerce weathers this more easily, because a large product catalogue gives the model far more to anchor on than a lead gen account ever has. A local emergency plumber has none of that ballast, so it drifts, and it drifts expensively. 

For accounts that have been treating keywords as thematic anchors and conversion data as the primary control mechanism, this transition lands a lot more softly, because that’s roughly how the system has been operating all along. The match types were already advisory.

I think the shift looks like this: Keyword selection becomes about intent representation rather than match-type containment; you’re choosing themes that accurately describe what your business does and who it serves at the moment of commercial intent. Negative keyword work gets more important because, without match-type walls, you’re the one actively managing the perimeter now, but it’s also going to be a very fine line between over-pruning and starving the system of the information it needs (yay). Conversion tracking integrity becomes the single most load-bearing element in the account because every expansion decision the model makes is downstream of whatever signal you feed it. And for lead gen specifically, that means interrogating what you’re even calling a conversion, because a form submission and a qualified lead are not the same thing, and the model can’t tell the difference unless you teach it.

Google reported a typical 14% lift in conversions or conversion value at a similar CPA for accounts that switch on AI Max, rising to 27% for non-retail campaigns that were still getting more than 70% of their conversions from exact and phrase match keywords. Those are Google’s own numbers, measured against campaigns without AI Max enabled, not against well-optimized accounts with clean signal architecture, so treat them as directional rather than a promise. The lift is real in some accounts. In others, what the model optimizes for isn’t what you actually wanted, and conversion volume climbs while lead quality quietly drops.

All of that turns on signal quality, and the piece most likely to get overlooked is the landing page. On a recent episode of Google’s Ads Decoded podcast, the recommendation was to run several RSAs in a single ad group, each pointing to its own landing page, and let AI Max mix and match across them through text customization and Final URL Expansion. Half of that holds up. Google’s own data cited on the episode put advertisers using text customization and Final URL Expansion at roughly 7% more conversions at a similar CPA, and Fred Vallaeys and the team at Optmyzr have seen the same shape in their RSA work, where ad groups with more asset variety earn higher impression share and stronger performance. A wider pool of creative for the system to mix from genuinely helps.

The half without data behind it is the landing pages. Adding another RSA is creative variety; adding another landing page is structural variety, and Final URL Expansion now lets the system decide which page a given search lands on. If those pages aren’t selling the same thing to the same person, the model gains another way to drift: toward whichever page-and-headline combination wins on engagement rather than the one that wins on commercial fit. You feel it downstream, in form fills that don’t close and sales calls where the prospect expected something you don’t sell. So before you add the second page, the question worth sitting with is whether both pages serve the same intent. The landing page is a signal like any other, and Final URL Expansion hands the system the same latitude over where your traffic goes that it now has over which searches you show for.

 

The thing worth paying attention to

Match types served us well for a long time. They gave us something that felt like control, and in the early years of paid search, they actually were control. The system matched what you typed, or very close to it, and you could build clean logic around that.

Unfortunately, my friends, that era is over. The matching logic has moved to a model that operates on semantic intent clusters, and no amount of careful keyword setup can put the wall back up. What you have instead is influence: over the signals you feed the system, the themes you anchor your keywords to, the conversion events you define as success, and the negative lists that keep the expansion territory honest.

That’s a different kind of discipline than match-type management. It asks you to think about your account as a signal architecture rather than a keyword list. The accounts that make that move cleanly, with tight conversion data, accurate keyword themes, and active perimeter management, will find that the expanded matching actually works in their favour. The accounts that treat this as a settings change rather than a strategic shift will fund the model’s confusion for months before they work out what happened.

The keyword survives because it’s one of the clearest signals in the system: the actual expression of what a person wants. The match type was the control layer we built on top of that signal, and the control layer is the part that aged out. Get the signal right, and the keyword outlasts every match type Google ever sold you.