AI Max search term matching: what it does to negatives

What actually changes on September 1

Google has been moving Search campaigns onto AI Max in stages since it reached general availability on April 15, 2026. The next stage has a date on it.

From September 1, 2026, two groups of Search campaigns are upgraded without anyone at the account doing anything:

• campaigns using automatically created assets are upgraded with search term matching and text customization switched on by default;
• campaigns using the campaign-level broad match setting are upgraded with search term matching only.

Two details decide how much this touches you. Final URL expansion follows text customization, so it does not arrive by default in the broad match group. And search term matching can be switched off at ad group level after the upgrade, which is a pass through your ad groups rather than one account switch.

Two dates around it are worth keeping straight. Creating new campaigns on the legacy setups ended on August 3, 2026, and the notice went out by email on August 5 rather than as a public announcement, which is why a lot of accounts have not heard about it. Dynamic Search Ads are not in this wave: that migration moved to February 2027 in a decision announced on June 11, 2026. If you were bracing for the DSA change, this is a different one, and it may hit campaigns you were not watching.

The reason this matters for search term work is narrow and specific. Search term matching changes which queries your ads become eligible for. Everything downstream of that, including what lands in your search terms report and what you have to review, changes with it.

Search term matching is two different things

Google describes search term matching as broad match plus keywordless technology. Those are not two names for one behavior. The Google Ads API reports them separately, and the enum values say what each one is:

Source	What it means
AI_MAX_BROAD_MATCH	Matches that expanded upon your existing keywords using AI-driven broad match logic
AI_MAX_KEYWORDLESS	Matches derived from your website content or landing pages, requiring no keyword

The first one is a wider version of something you already do. You wrote a keyword, Google stretched it. If you have run broad match, you have reviewed this kind of traffic before and you have a feel for how far it goes.

The second one is new to most accounts. No keyword is involved. Google reads your site and decides a query is close enough to what it found there. Nobody at the account asked for that query, no keyword report will explain it, and the query will not appear anywhere in your keyword structure because it does not belong to a keyword.

In the interface, incremental terms show up in the search terms report with an AI Max match type and a source column that separates the two. That column is the useful part of this whole change, and it is worth learning to read before September rather than after.

Do negative keywords still work

This is the question every account asks first, and it deserves a straight answer about what is documented and what is not.

Negative keywords still exist and still apply to Search campaigns. What is worth knowing is how Google frames them now. The help article on negative keywords, filed under search targeting and controls for its AI-powered campaigns, says to "use negative keywords only for essential brand safety needs or to block completely irrelevant terms," and calls the control "highly restrictive," warning that it "can harm performance by preventing the AI from finding valuable traffic." For your own brand it recommends brand exclusions instead, because those also catch misspellings, variants and subsidiary brands.

Meanwhile the help page describing how AI Max for Search campaigns works lists the campaign's controls as brand settings, locations of interest, URL exclusions and URL inclusions. Negative keywords are not named among them.

You will find blog posts stating that negative keywords are enforced before AI Max expansion, so that any query containing one is excluded regardless of what the model infers. That would be a useful guarantee. We have not found it in Google's own documentation, so we are not going to repeat it as a fact. What we can say is that the control is still there, Google's own guidance is to reach for it sparingly, and the safe assumption for planning is that your negative list is a filter you should verify on your own account rather than a rule you can take on trust.

There is a practical consequence either way. If you have been using negative keywords to keep your brand traffic in the brand campaign, check whether brand exclusions do that job better under AI Max. Those are different mechanisms, and the second one is the one Google points at.

Keywordless is the part to watch

If you only change one habit before September, make it this one: look at keywordless terms separately from everything else.

Adalysis published a useful piece on AI Max search term reporting on December 2, 2025 that lands on the same problem from the reporting side. Their analysis found AI Max search terms "without associated keywords" that could not be correlated to any keyword in the campaign, with "no way to find out why we're showing for these search terms." They also make a measurement point that is easy to get wrong: AI Max totals are not incremental gains, because expanded matches can include traffic you were already getting from exact and phrase keywords. Reading the AI Max row as pure upside overstates it.

A search term with no keyword next to it used to mean one thing: attribution was missing and you were looking at a gap in the data. On an AI Max campaign it means something else. It is an answer, not an absence, and the only lever over it is a negative keyword or a change to the site content the match was derived from.

Reading the source column without breaking your own numbers

There is a trap in pulling this data yourself, and we walked into it while building support for it, so here are the measurements.

segments.search_term_match_source is a GAQL segment, and a segment is a group-by. Adding it to your main search term query splits each term into one row per source and splits its metrics along with it. On one live AI Max campaign, the same query returned 599 rows without the segment and 614 rows with it, with identical impression, click and cost totals. 17 of 594 terms carried two sources. If you merge those rows back into one record per term, whichever row lands last wins and the term's recorded cost quietly becomes a fraction of what it really was.

The second measurement is about which rows exist at all. A Google Ads report only returns rows where at least one selected metric is non-zero, so the metrics in your SELECT decide your result set. On this exact query, clicks and cost alone returned 51 rows, every one of them with a click. Adding impressions returned 614. Most search terms never get a click, and keywordless matches are exactly the kind that tend not to.

The third is a reason not to run this everywhere. On a plain Search campaign the segment is not interesting: 334 of 334 rows on a live account came back as advertiser-provided keyword. Query it only for campaigns that have AI Max on, or you are paying for a second full scan of search_term_view to learn nothing.

If you are pulling this into a spreadsheet, the same rules apply. Keep the segmented pull as a separate tab, join it to your main report on purpose, and never let it overwrite your cost column.

What the evidence says so far

Google's numbers and independent numbers disagree, and both are worth having in front of you.

Google's case: at general availability on April 15, 2026 it cited an average improvement of 7% across the feature set, following a beta claim of 14% more conversions in May 2025.

The other side: Lunio published a study on invalid traffic on August 12, 2026 covering more than 414 million clicks tracked between October 2025 and June 2026. On retail search campaigns with AI Max enabled, its invalid traffic rate reached 5.28% in Q2 2026 against 3.07% on standard search campaigns in the same accounts, and AI Max accounted for 68% of all invalid clicks in its Google search dataset. Lunio sells invalid traffic protection, so read the framing accordingly, but the comparison is within the same accounts, which is the part that makes it hard to wave away.

Neither number tells you what will happen in your account. They tell you the range is wide enough that measuring is not optional.

A checklist for the two weeks either side

1. Find out which of your campaigns are in each group. Automatically created assets and the campaign-level broad match setting are separate lists and they get different settings.
2. Write down your baseline before September 1: cost, conversions, cost per conversion and search term count per campaign, for a period you will be able to compare against.
3. Decide ad group by ad group whether search term matching stays on. The opt-out is at that level, so treat it as a review, not a switch.
4. Check your brand handling separately. If negative keywords were doing brand containment, look at brand exclusions for that job.
5. Raise your search term review frequency for the first two to four weeks after the upgrade. This is the window where new query patterns show up and where a small negative list fixes what a large one would later have to undo.
6. Review keywordless terms as their own list, not mixed into the rest. They behave differently and they are where unrequested spend tends to collect.
7. Do not fan out broad negatives under pressure. A negative added in week one on a hunch is a negative nobody will dare remove in month six.
8. Re-check the numbers you baselined at 30 days, and de-duplicate AI Max terms against your exact and phrase terms before calling anything incremental.

Where NegativeShield fits, and where it does not

We built for this split before the September date was announced, because the API told us it was coming.

When a campaign has AI Max on, we run the match source as its own pass, for that campaign only, and store the sources on the term. A term Google matched against your site with no keyword shows up in the terms table marked "AI Max, no keyword" instead of an empty keyword cell, with the plain explanation: Google matched this query against your website, not against any keyword you added. Every verdict carries a one-line reason you can put in front of a client, nothing is pushed to the account until you approve it, and approved negatives are applied through the official Google Ads API as exclusion lists and shared sets.

What we do not do: we do not manage bids, budgets or ad text, we will not turn AI Max off for you, and we are not the fix for a campaign whose real problem is its offer or its landing page. If one person comfortably reviews one low-volume account, you probably do not need us for this. If what you want is a tool that runs the whole account on its own, that is a different category, and we compared the two models in NegativeShield vs Ryze AI.

If you want to see how the classification handles your own terms before connecting anything, the free CSV checker reviews up to 250 search terms. For a live account, request a free 20-minute audit call. If your review process is not written down yet, start with the search term audit workflow, and if you are doing this across a book of clients, the cross-account review queue is the piece that keeps September from turning into a fire drill.