Within a fortnight this summer, both of the major search advertising platforms acted on the same conclusion, which is that a keyword list maintained by a human is a lossy way to describe what people actually want. Microsoft rolled its AI Max product out across all accounts globally on the nineteenth of August. Google began automatically upgrading Search campaigns to a product of the same name on the first of September, working through the month. The shared name is an unhelpful coincidence and the two products differ in their detail, but the underlying judgement is identical and it represents the end of a twenty-year arrangement in which the advertiser told the auction what to match and the auction obliged.
What separates the two companies is not the technology. It is whether they asked.
Google Migrated Campaigns Without A Campaign-Level Opt-Out
Google moved every Search campaign using campaign-level broad match or legacy automatically created assets onto the new system during September, with no opt-out available at campaign level, though individual features within it remained controllable in campaign settings. The mechanics are worth understanding precisely, because much of the coverage described this as a new campaign type and it is nothing of the kind. AI Max is an optimisation layer activated inside existing Search campaigns, which means campaign identifiers and account structures survive untouched. Three capabilities sit within it: search term matching, which reaches queries the keyword list does not cover; text customisation, which generates and selects messaging variants at auction time from existing assets and website content; and final URL expansion, which routes a visitor to whichever page best matches their intent rather than to a fixed landing page. Alongside the migration, Google closed the door on new Dynamic Search Ads campaigns on the third of August and scheduled that format for full retirement in February 2027, ending a product that has been part of search advertising for more than a decade.
Microsoft Shipped A Toggle And Three Checkboxes Instead
Microsoft reached the same technical conclusion and handled the transition in the opposite manner. Its global rollout, announced on the nineteenth of August by Navah Hopkins, the company's advertising liaison, arrived as an opt-in product rather than a migration. Advertisers see a toggle that can be switched off and three checkboxes that can be unticked individually, and the features do not activate unless somebody chooses them. Three controls shipped on day one rather than arriving later in response to complaints: brand controls allowing up to twenty brand lists per campaign split into inclusions and exclusions, term exclusions preventing specified phrases from entering AI-generated ad copy, and URL exclusions blocking particular pages from receiving paid traffic. Microsoft also made per-feature optimisation experiments available, so an advertiser can test individual components separately rather than accepting or rejecting the bundle.
The contrast is instructive because Google eventually built similar controls, just in the other order. Branded search controls inside Google's product were first spotted on the twenty-ninth of May and rolled out from the first of June, offering three settings: show ads on all relevant searches, which remains the default, control branded searches through inclusions and exclusions, or show ads only on unbranded searches. The problem they address is the pulling of branded queries into prospecting campaigns, which raises the cost of clicks a brand would have won regardless and blurs the boundary between branded and non-branded performance. Manual negative keyword lists had been the workaround and they consistently leaked on misspellings and variants. Both companies arrived at roughly the same control set. One of them had it in place before asking advertisers to trust the system, and the other did not.
Microsoft's Eight Percent Comes From Ten Of Forty-Four Experiments
Microsoft published the more transparent numbers of the two, and its methodology deserves credit before its headline gets scrutiny. The underlying work covered forty-four advertiser-run A/B experiments conducted globally between June and August, comparing Search campaigns with the feature enabled against the same campaigns with it disabled while holding bid strategy and budget constant, which removes the two variables that usually contaminate this kind of testing. Across all forty-four, the spend-weighted conversion uplift came out at approximately 13.6 percent.
The figure that travelled, however, was "at least 8 percent," and its provenance is narrower than that phrasing suggests. Ten of the forty-four experiments reached statistical significance, and eight percent is the lower bound of the confidence interval drawn from that subset, rounded down. The methodology sits in a footnote to a post published on the twenty-seventh of August by Carolyn Chou, a director of product management, rather than in the newsletter that carried the claim, and the outcomes of the remaining thirty-four experiments have not been described, so whether they were positive but underpowered, flat, or negative is not public. None of this makes the result wrong, and ten significant outcomes from forty-four underpowered tests is an unremarkable thing to find. It does mean the number is more specific than its circulation implies. Microsoft also reported one finding that holds up well against independent work, which is that uplift concentrated in accounts leaning heavily on exact and phrase match keywords, and BBQGuys, the named advertiser in the release, described incremental conversions at higher average order value from long-tail queries its manual targeting had never reached.
Google And Its Sharpest Critic Both Land Near Thirteen Percent
Google puts its own uplift figure at fourteen percent, and separately reports that campaigns using the full feature suite see around seven percent more conversions or conversion value at similar cost per acquisition compared with search term matching alone, drawn from hundreds of thousands of advertisers. Set against that, Smarter Ecommerce, an independent firm with a track record of scepticism about platform automation, analysed more than two hundred and fifty Search campaigns and around a million impressions and found median conversion value up thirteen percent.
Those two numbers landing within a point of each other is the most useful thing in this story, because in an industry where vendor figures and third-party figures routinely diverge by an order of magnitude, a platform and its most pointed critic agreeing on the size of the revenue effect ought to settle the question of whether the volume gain is real. It is. Where the two accounts part company is on what that volume costs. Smarter Ecommerce found median cost per acquisition rising sixteen percent alongside the thirteen percent revenue gain, with return on ad spend outcomes spread from forty-two percent above baseline to thirty-five percent below it, and only twenty-two percent of campaigns landing anywhere near their original targets. Mike Ryan, who led the analysis as the firm's head of ecommerce insights, summarised it without hedging: turning the feature on is essentially a coin toss, and a lift may arrive without efficiency following it.
Exact Match Keywords Are Now About Seventy-One Percent Exact
The finding most likely to surprise anyone running search advertising concerns how much of this had already happened before either announcement. A separate analysis by the same firm tracked 383 million impressions across EMEA ecommerce Search campaigns between January 2025 and July 2026, measuring how often an impression on an exact match keyword actually resolved as a genuine exact match. At the start of that window the answer was essentially always. By July 2026 it had fallen to roughly seventy-one percent, with about twenty-nine percent of exact match impressions being expanded by AI-driven search term matching, and around two-thirds of that entire movement occurring in the four months between March and July.
The composition of those expanded impressions is more surprising still. Of the AI-expanded impressions measured, 80.11 percent originated from exact match keywords and 19.52 percent from phrase match, with broad match contributing 0.38 percent. The expansion is concentrated almost entirely at the tightly controlled end of accounts, which is precisely where advertisers had assumed their settings were holding and where most teams place their highest-intent spend.
One Account Sent 69% Of Its Expanded Impressions To Competitor Terms
The mechanisms producing the efficiency spread turn out to be specific rather than mysterious, and two of them are striking enough to check for directly. In some accounts examined, as much as sixty-three percent of the expanded coverage was recycling queries the account was already winning through its existing keywords, which produces the appearance of incremental reach while delivering none of it. In one account, sixty-nine percent of expanded impressions landed on competitor terms, a placement most advertisers would choose deliberately or not at all rather than arrive at by default. Compounding both, nearly half the accounts in the study were running AI-expanded Search alongside Dynamic Search Ads and Performance Max simultaneously against overlapping inventory, which fragments attribution before anyone attempts to read a result.
Google did expose new reporting surfaces alongside the migration, separating keyword-driven traffic from AI-expanded traffic, showing which search term, landing page and headline were served together, and revealing the URL that was actually served. That last one exists because final URL expansion can route a visitor somewhere other than the configured page, which silently breaks any reporting pipeline that joins on ad URL and surfaces weeks later as unexplained data quality noise rather than an obvious error.
Ready to make AI part of your marketing strategy? Let’s talk.
Google's Published Figures Still Exclude Retail
Three questions remain open, and each should be answered within a couple of quarters. The first is whether Google publishes performance data that includes retail, since Search Engine Land noted that its fourteen percent figure excludes the vertical where the independent evidence is least flattering, and closing that gap would resolve much of the present argument. The second is whether Microsoft characterises the thirty-four experiments it has left undescribed, which would strengthen rather than weaken its position given that underpowered tests routinely fail to reach significance without indicating anything is wrong. The third is whether the efficiency spread narrows. If the range from forty-two percent above target to thirty-five percent below is mostly a function of account structure, brand separation and reporting hygiene rather than the technology itself, it should compress as those practices become standard. If it is still that wide going into 2027, the argument that this traffic is genuinely incremental becomes considerably harder to sustain.
What is not in question is that the change has happened. Both major search platforms have moved AI-driven matching from an option to the centre of how their auctions work, and the evidence assembled so far says the additional volume is real while the cost of acquiring it varies far more than any advertiser would like. The open question is whether the industry notices that one company asked permission first.
