AI Max for Search does something to your account that the match-type column does not show: it stops honouring the match types you set. Keywords you deliberately made exact or phrase, to hold them to a narrow set of queries, are matched far more loosely once AI Max is on, with its query-expansion layer deciding what counts as close enough. Search Engine Land's teardown of the feature put the harm plainly — AI Max undermines match-type control, so an exact brand keyword can begin serving on non-brand and competitor queries while a generic keyword starts matching a rival's brand name. The control you relied on to keep spend where you aimed it is quietly gone, and nothing in the keyword table tells you so.
That is the whole problem this post is about: not that AI Max's reporting is blurry — other posts cover that — but that your deliberate match-type targeting has been overridden, and the damage shows up as brand-term bleed you can only see in the search terms report. The fix is a two-step job you can do this week: detect the bleed by reading which queries your tight keywords are actually catching, then fence it with the right combination of brand exclusions and negative keywords — without over-cutting and throwing away the converting reach AI Max exists to find. Read the queries first; choose the lever second.
What match-type collapse actually means
Match-type collapse is the gap between the match type shown against a keyword and the match type that governs its serving. Normally an exact-match keyword only triggers on that query and its close variants, a phrase-match keyword on queries containing the phrase, and broad on the widest set; you pick among them to trade reach against control. Under AI Max that trade no longer holds, because the expansion layer sits in front of every keyword and matches on intent rather than on the match-type rule. The keyword still reads “exact” in the interface, but it behaves as something much closer to broad, and the queries it catches are chosen by a model, not by the operator who typed the keyword.
This is a different failure from the one most AI Max coverage focuses on. The reporting problems — hidden terms, AI-interpreted queries, attribution that does not reconcile — are about not being able to see clearly. Match-type collapse is about losing a control you were actively using: you set exact precisely so a keyword would not stray, and the setting is now cosmetic. That matters for anyone who structured an account around match type to separate brand from non-brand, protect a tight converter, or keep a broad experiment walled off from the rest. If exact no longer means exact, every one of those structural decisions is leaking at the edges, and the first place the leak surfaces is the search terms report, not the keyword report where you set the type.
How brand-term bleed shows up in the search terms report
You detect the bleed by reading the search terms report at the keyword level, not by scanning totals. Filter to your brand campaign or brand keywords and look at the actual queries they served on. If the account is clean, the queries under an exact brand keyword are your brand name and its obvious close variants. Under AI Max you will often find something else: generic non-brand searches, category terms, and — the expensive case — competitor brand names, all attributed to a keyword you set to exact so it would catch none of those. Each of those rows is a search your brand keyword should never have matched, and each one is money spent at brand-campaign bids on audiences that were not looking for you.
The bleed runs the other way too, and it is easy to miss because you are looking in the wrong campaign for it. Your generic non-brand keywords, expanded by AI Max, can start matching competitor brand queries, so you end up advertising against a rival's name without ever having added it as a keyword. Adalysis, testing the feature, reported exactly this bidirectional pattern — brand terms matching non-brand queries and non-brand terms matching competitors — in its analysis of AI Max search-term reporting. To find it, sort the non-brand campaign's search terms by cost and scan for brand-shaped queries you do not own. The tell in both directions is the same: a query whose audience does not match the keyword's intended audience, surfacing only because expansion put it there. Nothing above the search-terms level will show you this, which is why the detection step is non-negotiable before you touch any control.
Fence it: brand exclusions and negatives do different jobs
Once you have the bleeding queries in front of you, the fix is to fence them — and the right fence depends on what the leak is. Brand exclusions are the efficient tool for a whole competitor name: a brand list applied at the campaign level stops your ads serving on searches Google classifies as belonging to the brands you list, so one entry fences off a rival's name and the variants Google recognises for it in a single move. That is the lever to reach for when your generic keywords are bleeding onto a competitor's brand, because enumerating every form of a brand name by hand is a losing game and the brand list does it by classification instead.
Negative keywords do the job brand exclusions cannot. They stop specific queries or fragments regardless of how Google classifies the brand, which is how you catch the misspellings, the brand-plus-modifier searches, and the generic phrases that a brand list misses because Google never recognised them as the brand in the first place. The honest limitation, which Google's own documentation is clear about, is that neither control is complete on its own: brand matching leaves gaps on unrecognised variants, and negatives mean listing every variant yourself. The two together are how you get coverage without infinite manual work — the brand list for the bulk of a competitor's name, negatives for the specific leaks it does not catch. This division of labour is the same one covered in depth in brand inclusions, exclusions and negatives and, for Performance Max specifically, in PMax brand exclusions versus negatives; the new wrinkle under AI Max is only that you now need them to repair match-type control you used to get for free.
Add your tight keywords as broad to keep the bleed measurable
There is a counterintuitive instrumentation step that makes all of this visible and keeps it visible. Add your existing exact and phrase keywords to the campaign a second time as broad match terms. Because AI Max already matches everything loosely, this does not widen your targeting — the expansion was happening regardless — but it gives the report a labelled broad-match version of each term to attribute traffic to, so you can once again tell which queries came from your tight keywords and which came from expansion. Without that label, AI Max tends to credit the loosely matched traffic to your exact and phrase keywords, flattering them and hiding how much of their apparent performance is really expansion you did not ask for.
This is the same remedy Adalysis recommends, and it is worth being clear about why it is a diagnostic rather than a concession. You are not loosening control you still have; you already lost that control the moment AI Max went on. What you are buying back is measurement — the ability to separate the keyword's own earned traffic from the model's expansion so you can judge each on its merits. Once the account is instrumented this way, the bleed stops being invisible: expansion traffic lands against the labelled broad terms, your tight keywords show what they genuinely caught, and the brand-shaped queries that should not be there stand out instead of hiding inside an exact keyword's totals. Detection and this instrumentation reinforce each other — the labels make each subsequent search-terms pass faster and more honest.
Do not over-fence: keep the reach AI Max is for
The failure mode on the other side is fencing so aggressively that you defeat the reason you ran AI Max. The feature exists to find converting queries you would not have targeted yourself, and some of the expanded traffic — including some that arrived through match-type collapse — will genuinely convert at an acceptable cost. If you react to the bleed by blanket-excluding every non-exact match, you throw that away along with the waste, and you would have been better off not enabling AI Max at all. The goal is not to restore exact-match purity; it is to stop the specific leaks that are clearly off-audience while leaving the profitable surprises alone.
So apply the same significance discipline you would to any negatives decision. A query bleeding onto a competitor's brand name, or an exact brand keyword serving plainly irrelevant generic searches, is a clear cut — fence it. A non-brand query that arrived through expansion and is converting at a cost you would accept is reach you keep, whatever match type notionally produced it. For the borderline rows, let them accumulate enough spend to be a signal before you act, exactly as argued in how many clicks before negating a search term. The point of detecting the bleed first is precisely so you fence by evidence rather than by reflex; the collapse of match-type control is a reason to watch the queries more closely, not a reason to amputate every query a tight keyword did not strictly own.
Fit it into the audit routine
Make the bleed check a standing part of your AI Max search-terms pass rather than a one-off panic after you read about the feature. Each cycle, filter the search terms report to your brand keywords and confirm they are still catching brand queries and nothing else, then sort the non-brand campaigns by cost and scan for brand-shaped queries that should not be there. Where you find bleed, route it: a whole competitor name to the brand exclusion list, specific variants and generic leaks to negatives, and keep your tight keywords mirrored as labelled broad terms so the next pass is faster. This slots directly into the broader search hygiene audit playbook as the match-type integrity step.
Two related decisions sit next to this one and are worth reading alongside it. When a competitor query is genuinely converting, the question of whether to fence it or bid on it deliberately is covered in competitor search terms: negate or bid — brand bleed is sometimes an accidental version of a bid you might choose to make on purpose. And because AI Max's reported numbers will not line up with what it actually served, reconciling the two is a separate skill, handled in AI Max attribution discrepancies. Together those three — detect-and-fence the match-type collapse here, decide bid-versus-negate on competitor terms, and reconcile the attribution — are the core of keeping an AI Max account honest once the match types you set stopped meaning what they say.