Google just let the tight keywords into AI Mode

What actually changed this week

The fox does not force the fence, it finds the gap somebody left. Advertisers were told for a year that reaching AI Mode meant handing targeting to automation, and this week a small door opened in that fence.

Read nextGoogle Ads AI Max arrives on 1 September whether you asked or not/ChatGPT Ads vs Google Search Ads: Which Should You Budget for?

Google is running what it calls a small experiment in which Search campaigns using exact and phrase match keywords can serve text ads in AI Mode, a placement previously reserved for AI Max and Performance Max. Ads Liaison Ginny Marvin confirmed the test after AdSquire founder Anthony Higman spotted it, as Search Engine Land reported on 7 September.

There is a condition attached, and it is the whole story. Ads qualify only where Google identifies explicit and direct user intent, so tight keywords will not automatically reach every AI Mode placement. The gate is open. It is not wide.

Why this matters more than a feature note

Until now the trade was explicit: to appear where conversational queries land, you accepted looser matching and less control over the query set. That trade fell hardest on regulated advertisers, who cannot let an automated system wander into a query they are not licensed to answer.

If tight keywords can reach the same surface, the compliance argument for staying on restrictive match types stops costing reach. That is a meaningful change for anyone whose ad copy has to survive a regulator reading it.

Where the queries actually went

Before rebuilding anything, look at where the demand is sitting. Two independent measurements from this summer draw the same shape from different directions.

Those figures come from Conductor's 2026 Publishing AI Search benchmarks, which also found that 57.8 percent of Publishing AI Overview citations came from pages outside Google's organic top ten. Position and presence have come apart, and the paid layer is now following the same logic.

The traffic picture points the same way. Using Semrush data, Search Engine Land compared July 2026 with July 2025 and found ChatGPT up 48.38 percent to 1.09 billion monthly visits, YouTube up 36.6 percent, and Google itself up 10.72 percent to 25.3 billion. Bing fell 50.43 percent over the same period.

Read those two charts together rather than separately. Search is not shrinking, it is answering differently, and the destination people reach for is increasingly a conversation rather than a list.

What match types mean now

Time for the unglamorous refresher, because half the arguments about this test are really arguments about definitions.

Google's own documentation is clear that broad match is the default match type and reaches every search that exact and phrase match reach, plus more. It is also the only match type that uses the full set of signals available in the account, which is the answer to the question everyone asks: what signals does broad match use, and the published list includes landing pages, the other keywords in the ad group and the user's previous searches, per Google's guide to broad match.

The broad match vs phrase match question therefore has a new axis. It used to be a trade between reach and control. Now it is a trade between reach, control and which surface you are allowed to appear on, and that third term was not in anybody's spreadsheet a fortnight ago.

The modifier that keeps coming back

Every discussion of this reaches for the broad match modifier, which Google retired years ago and which practitioners still search for constantly. That search demand is a tell: buyers want a middle setting between meaning-matched and query-expanded, and no such setting exists today.

The nearest honest substitute is a disciplined structure. Tight keywords in campaigns you watch closely, broad match keywords in a separate campaign with its own budget and its own negative list, and a weekly read of the search terms report to move phrases between them. Dull, effective, and it survives whatever this test becomes.

What broad match documentation actually says

The published definitions have moved more than most accounts have. Google's keyword matching overview and its match options page both now describe matching as a meaning problem rather than a string problem, and the campaign-level broad match setting documents how the loosest option is applied across a whole campaign rather than keyword by keyword.

That last page is the one to read twice, because it is the mechanism that carried accounts into automated expansion in the first place. A campaign-level switch is quiet in a way a keyword change is not, and quiet settings are where budgets wander off into the undergrowth.

On the organic side, Google's AI features documentation still says no special markup or file is required to appear in its generative surfaces. Anyone selling a technical shortcut into the answer layer is selling a scent with no fox at the end of it, and the paid experiment reported this week does not change that.

A second signal from the shopping side

Something adjacent is happening in Shopping, and it is worth reading beside this because it shows how carefully these effects need to be described.

Mike Ryan, head of ecommerce at Smarter Ecommerce, looked across thousands of Shopping and Performance Max campaigns spanning hundreds of advertiser accounts and found a consistent pattern: clicks relatively flat or slightly down, while impressions fell more noticeably. Optmyzr's own data showed click-through rate up 17 percent year over year, close to the roughly 20 percent rise Ryan was seeing, per Search Engine Land.

Here is the part that matters, in his own words: the data does not prove AI Overviews are causing the change, the timing could be coincidental, and something else entirely could be responsible. He calls it a hypothesis. That is what an honest read of a moving system sounds like.

A rising click-through rate that comes from losing your weakest impressions is not the same achievement as a rising click-through rate that comes from better ads. Both look identical in a dashboard. Only one of them is worth celebrating, and telling them apart requires impressions in the same view.

Five moves worth making this week

None of these depend on the experiment surviving. They all pay for themselves if it does.

For regulated advertisers there is a sixth, which is to document the query set you are willing to appear against before you widen anything. That document is the difference between a compliance conversation and a compliance incident.

The wider lesson holds regardless of what Google does with this test. Placement is being decided by intent classification rather than by match type alone, and the accounts that will cope are the ones whose reporting can already tell one surface from another.

If you want that reporting built and the match-type structure rebuilt behind it, that is folkfox PPC, and it sits next to the organic side in SEO and GEO. We wrote about the other half of this shift when AI Max arrived whether advertisers asked or not, and about the budget question in ChatGPT Ads against Google Search Ads. The full stream is on folkfox news.

A closing thought on posture. The fox does not sprint at the first rustle in the hedgerow, it waits to see which way the grass moves. Read the test, segment the reporting, keep the trail of what changed, and let the experiment prove itself in your own data before it rearranges your account.

If it holds, the reward is real: reach into the answer layer without surrendering the query set, which is the trade regulated advertisers have wanted since AI Mode arrived. If it does not hold, you will have spent a fortnight improving your reporting, which was overdue anyway.

Frequently asked questions