On September 4, 2026, Google Ads Liaison Ginny Marvin confirmed to Search Engine Roundtable that exact and phrase match keywords in ordinary Search campaigns are now eligible to serve text ads inside AI Mode. Until that confirmation, the only routes into AI Mode ad inventory were AI Max, Performance Max for Search, or broad match paired with Smart Bidding.
Nobody opted in. There is no checkbox to opt out and no reporting segment that tells you it happened. For Michigan advertisers running tight, keyword-controlled accounts, that combination is the story.
What changed in Google Ads on September 4, 2026?
Exact and phrase match keywords in standard Search campaigns can now trigger text ads on AI Mode results pages. Google is calling it a limited experiment, and it applies without any account-level setting change.
Anthony Higman spotted it first, posting on LinkedIn that “Google is now serving regular pure search campaigns with restrictive match types in AI mode,” confirmed “through multiple campaigns and tests on our side.” Marvin’s reply to Search Engine Roundtable was short: “This is a small experiment that recently kicked off.”
Her fuller explanation in the LinkedIn comments is worth reading twice. Exact and phrase match keywords are eligible to serve text ads in AI Mode, she wrote, but “this is limited to cases where there is explicit and direct user intent.” AI Max and PMax for Search, she added, “are designed to capture complex conversational intent and access our latest ad formats (like Highlighted Answers).”
Two tiers, then: standard Search gets the narrow literal slice of AI Mode; automated campaign types keep the conversational slice and the newer ad units.
Why does this matter if you never turned on AI Mode?
Because it moves your money to a different surface without changing anything visible in the interface. Standard Search campaigns have no AI Mode network segment; those impressions fold into the same Google Search row you already report on.
Consider an industrial distributor in Livonia bidding exact match on a specific valve size and thread spec. That bid was priced against a results page where the buyer still had to click something to learn anything. The same keyword may now serve beside a generated answer that has already told the buyer which part to order and why. The click can cost the same and convert at half the rate, or at twice the rate because the buyer arrives pre-qualified. Both are plausible. Neither is labeled.
The stakes scale down badly. Most Michigan SMB accounts run between $4,000 and $15,000 a month. At that size, shifting eight percent of impressions onto a surface with different click behavior does not announce itself — it surfaces six weeks later as cost-per-acquisition drift, and the team spends three of those weeks rewriting a landing page that was never the problem. That is the kind of unattributed variance a disciplined Google Ads management review catches before it compounds.
How does an exact match keyword end up on a conversational query?
Two classifiers now stack. Google first decides whether an AI Mode prompt carries explicit, direct commercial intent. Only prompts clearing that filter reach the auction, where close-variant matching then decides which keyword the prompt means.
Exact match has not meant “the exact words” since 2018. It means same meaning, resolved by close-variant matching — a system built for typed queries of two to five words. AI Mode prompts routinely run fifteen to thirty. The intent filter bridges the gap by discarding the long deliberative prompts before matching ever runs, leaving the short declarative ones that look structurally like traditional search queries.
Understand that and you can predict cases nobody has written about. An orthodontic practice in Rochester Hills bidding exact on a cost keyword will pick up AI Mode traffic, because “what does clear aligner treatment cost near me” is explicit and direct. The same practice bidding on a comparison keyword will not: “is clear aligner treatment worth it versus braces” is deliberative, and routes to the conversational tier only AI Max and Performance Max reach. The filter selects for grammatical directness, not topic.
Standard-campaign advertisers therefore inherit the narrowest and most expensive slice of AI Mode: bottom-of-funnel prompts. Expect average CPC on your tightest exact match keywords to move first.
Is “a small experiment” the right way to read this?
Marvin’s phrasing is accurate about volume and understated about consequence. Higman’s counter is sharper, and plainly made: before this, “the ONLY way to show in those places was through ai max, p-max or broad and smart bidding.”
That gate was doing real work. Inventory exclusivity — not persuasion — has been Google’s main lever for moving budget into automated campaign types. Opening AI Mode to exact and phrase match removes one of the last concrete reasons a skeptical advertiser had to migrate, which reads less like generosity than retention: advertisers refusing AI Max were paying an inventory penalty, and the penalty just shrank.
The PPC industry is treating this as a win, and that reaction deserves pushback. Reach you cannot segment, exclude, or attribute is not reach. It is variance dressed as opportunity. File this under measurement risk for September, not growth. If it becomes durable incremental volume, that will be visible by October, and you can celebrate then with numbers.
How should Michigan advertisers audit their accounts this week?
The goal is not to react to the experiment. It is to preserve a clean before-and-after.
- Export a baseline today. Pull the search terms report plus keyword-level impressions, average CPC, CTR, and conversion rate for August 21 through September 3, 2026. That two-week window is the only clean pre-experiment control you will get.
- Segment by match type, not by campaign. If your reporting rolls exact and phrase together with broad, split them now, or the signal disappears into the average.
- Flag any exact or phrase keyword whose average CPC or CTR moved more than 15 percent after September 4 with no bid, budget, or ad copy change behind it. Those are your candidates, not proof.
- Watch query length in the search terms report. A sudden cluster of eight-to-twelve-word queries against a tight exact keyword is the most legible fingerprint available.
- Read your landing page as if the answer already happened. If a generated answer resolved the question before the click, a page that restates it wastes the visit. Lead with price, availability, fitment confirmation, or a booking step in the first screen.
- Add campaign-level negatives for deliberative phrasing you do not want to fund — comparison and “worth it” modifiers — but only where conversion data already proves they lose money. Do not negate on theory.
- Hold your tCPA and tROAS targets for at least 14 days. Loosening targets to absorb an unexplained CPA move teaches Smart Bidding a lesson you cannot unwind cleanly.
- Name an owner and a recheck date. One person, September 19, thirty minutes. Anything vaguer does not survive a busy month.
What dates should be on the calendar right now?
Several dated changes overlap this experiment, which is what makes September hard to read cleanly.
- September 1–30, 2026 — Google’s automatic migration window for campaign-level broad match and automatically created assets campaigns into AI Max. If any of your campaigns migrate mid-month, that alone will move your numbers.
- September 1, 2026 — Google added a paragraph to the AI Max help docs on keyword prioritization: when you disable the broad match campaign setting or move to AI Max, broad keywords are not treated as exact match in prioritization. Preserving that behavior means adding copies of your exact match keywords, and converting broad to exact wipes historical keyword stats.
- September 4, 2026 — the exact and phrase match AI Mode experiment confirmed. Treat this as day zero for your before-and-after.
- Late September 2026 — the campaign-level language setting is removed from Search and AI Max for Search. Google will rely on creative language, landing page language, and its own model of what languages a user knows. Metro Detroit advertisers running Arabic, Spanish, or Bengali pages should confirm creative and page language agree.
Four changes inside thirty days puts single-variable attribution off the table for September. Document what you changed and when, or your October retrospective is guesswork. We keep a running log of these shifts for Michigan businesses precisely because they arrive stacked.
Frequently asked questions
Can I opt out of serving in AI Mode?
No. As of September 5, 2026, Google has not published an opt-out control for AI Mode serving in standard Search campaigns, and none appears in the campaign settings interface. Your practical levers are negative keywords, match type discipline, and bid adjustments. If you need to reduce exposure, tightening keyword coverage is the only reliable route.
Will I see AI Mode traffic broken out in my reports?
Not currently. There is no AI Mode network segment for Search campaigns, so these impressions report inside the standard Google Search row. Query length shifts in the search terms report are the closest thing to a signal available, and they are circumstantial rather than definitive. Build your baseline now so the comparison is possible later.
Does this mean I should stop planning an AI Max migration?
It weakens one argument, not the whole case. AI Max and Performance Max for Search still reach conversational AI Mode intent and still access formats like Highlighted Answers, which exact and phrase match do not. If your migration was motivated purely by inventory access, revisit it. If it was motivated by query coverage, the reasoning is unchanged.
How confident should I be that this is affecting my account?
Low, for now. Google described it as a small experiment, and experiments are typically limited by geography, vertical, and account sample. Higman reported confirmation across multiple campaigns, which is meaningful but not a scope statement. Treat any September CPC movement as unexplained until you have two clean weeks of post-September 4 data.
Export your baseline, split reporting by match type, hold your targets, recheck September 19. Under an hour of work, and the difference between knowing what happened to your account and arguing about it in November. The experiment may amount to nothing. The measurement gap it exposed will not.
Comments are closed