AI Max Negative Keywords: How PPC Agencies Control Search Term Spend

Export the AI Max search terms, have an AI sort them by intent, then approve negatives at the narrowest level that works: ad group, campaign, then account list. Protect any term that has converted before adding a negative, and repeat weekly on terms with meaningful spend. The lead approves classes of waste, not hundreds of individual terms.

Two facts change how negatives work under AI Max. Google's help pages say that turning AI Max on switches search term matching on in every ad group by default, using broad match, asset-based and landing-page-based matching, so your exact and phrase keywords no longer mark the edge of your reach, and part of the traffic arrives with no keyword behind it at all. Negatives and exclusions are therefore the main brake the agency still holds, and the person approving spend should know which ones are set.

Follow me on Instagram@sagnikteaches

What AI Max changes before you add a single negative

AI Max is not a new campaign type. Google's API documentation calls it an optimisation layer you switch on inside an existing Search campaign. It bundles search term matching, which uses broad match and keywordless technology to reach beyond your keywords, with text customisation (automatically generated headlines and descriptions) and final URL expansion, which can send people to a page you did not choose. The API documentation says enabling AI Max switches on only search term matching by default, and final URL expansion needs text customisation turned on, so check which of the three each client's campaign actually has on. For whether each of those settings suits a client, see should you use AI Max for Google Ads.

Connect on LinkedInSagnik Bhattacharya

Two details matter to whoever signs off spend. First, Google's help page on how AI Max works says search term matching is an ad group setting, on by default once AI Max is on for the campaign, and you can switch it off in the ad group settings. An account built on tight exact and phrase keywords is no longer tight on the day AI Max goes live, because its reach now runs past those keywords. Second, the AI Max reporting page shows how each query was matched. The API field is segments.search_term_match_source, and the AI Max values are AI_MAX_BROAD_MATCH (an expansion of your existing keywords) and AI_MAX_KEYWORDLESS (matches drawn from your website content or landing pages, with no keyword needed). In the interface, the search terms report shows "AI Max" as a match type, with a source column saying whether the match came from broad match expansion or keywordless matching.

Subscribe on YouTube@sagnikteaches

When Google launched AI Max in May 2025 it said advertisers who switch it on will typically see 14% more conversions or conversion value at a similar cost per conversion or return on spend, based on its own data, and a typical 27% for campaigns that relied mostly on exact and phrase keywords. Treat that as a claim to test on each client, not a forecast. The weekly routine below is how you test it, and the negative list is what keeps the test honest.

Where each control lives, and what it can and cannot block

Use the narrowest control that does the job. The table lists the six you have and the trap in each.

Google's help page on account-level negatives states the 1,000 limit (checked November 2026), and Google can change it, so confirm it in the account before planning a long list. For the account-level mechanics in a sister product, see Performance Max negative keywords: account-level controls.

The export: what to pull and what to leave out

Pull the search terms report with the match source shown, for the last 28 days the first time and the last 7 days each week after. Keep only terms with spend above zero, sorted by cost. These are the columns the AI needs, and nothing else:

• Search term, match source, campaign, ad group and landing page (the AI Max search term and ad combination view shows the combination of search term, headline and landing page).
• Impressions, clicks, cost, conversions and conversion value.
• The keyword, where one exists. Keywordless rows will have none.

Two cautions from Google's documentation save embarrassment later. It says never to sum metrics from the standard search term view and the AI Max search term and ad combination view, so pick one for each calculation. And campaign totals will not match the sum of the exported terms, since low-volume terms are left out of reports for privacy. Say so in the first report, or someone will think the export lost money.

Before the file goes anywhere, scrub it. Search terms are what people typed, so they can contain names, addresses or phone numbers. Drop the account name, remove obvious personal details, and check the client agreement and the data controls on the AI plan you use. Keep each upload to a few thousand rows by splitting per campaign, because a shorter table is easier to spot-check and less likely to hit the tool's limits.

A classification prompt that returns decisions, not opinions

The prompt below asks for one row per term with a class, a confidence, the shortest safe negative and a reason. It tells the model to say UNCLEAR rather than guess, and it hands over the converting terms so the AI can see what must not be touched.

You are helping a PPC analyst review search terms from a Google Ads Search campaign with AI Max switched on.

BUSINESS: [what the client sells, who buys, what they do not sell]
BUYER SIGNALS: [for example quote, price, cost, installer, installation, business, home]
TARGET COST PER CONVERSION: $[120]
PROTECTED TERMS: [paste every search term that converted in the last 90 days]

For each search term in the table below, return one row with:
1. CLASS: BUYER, RESEARCH, COMPETITOR, JOB SEEKER, FREE/DIY, WRONG PRODUCT, WRONG LOCATION or UNCLEAR
2. CONFIDENCE: high, medium or low
3. PROPOSED NEGATIVE: the shortest phrase that blocks this term without containing any buyer signal or protected term. Write NONE for BUYER, RESEARCH, COMPETITOR and UNCLEAR.
4. MATCH TYPE: exact or phrase (never broad)
5. REASON: under 12 words

Rules: never mark a term as junk if it contains a protected term or a buyer signal. If unsure, write UNCLEAR. Do not invent data. Return a table only.

An illustrative result for a security installer's campaign (28 days, target cost per conversion $120), showing 8 of 640 rows:

Three things in that output deserve a human eye. The AI correctly refused to propose a bare "free", for reasons the next section shows. It wrote "unclear" for the doorbell term, which is the right call: whether the client fits and supports that product is a question for the client, not for a model. And it left "best cctv brands" alone because the term has not yet earned a block: $172 of spend is under the $240 threshold, so it goes on the watch list. Each week, spot-check 20 rows from the junk pile and 20 from the buyer pile. If more than two in either are wrong, fix the prompt before approving anything.

The protect check: keep negatives away from buyers

A phrase negative blocks every query that contains those words in that order, whatever else surrounds them. That is the hazard. A bare "free" negative would have blocked "cctv installation free quote", which cost $212 and produced 6 conversions in the sample, the best-performing row on the list. The protect check catches this mechanically, before anyone approves anything.

Put every search term that converted in the last 90 days on one sheet called Converting. Put the proposed negatives in column A of another. In column B, count how many converting terms contain each proposed negative:

=SUMPRODUCT(--ISNUMBER(SEARCH(A2, Converting!A:A)))

Any row with a result above zero is held and either narrowed ("free cctv" instead of "free") or dropped. The check works for phrase and exact negatives, and that is why the prompt forbids broad ones: a broad negative blocks any query containing all its words in any order, which needs a stricter test than a simple substring search. Also hold anything that contains one of the client's service names combined with a buyer signal such as quote, price or installer, even if that exact query has not converted yet. In the first clean-up below, this check removed 9 proposals.

Negatives also do not match close variants the way positive keywords do. Google's help page on negative keywords says you have to add synonyms and singular or plural versions yourself, while casing and misspellings are accounted for automatically. So a "jobs" negative does not block "cctv installer job": add "job" as well, and do the same for every word on the list that has a singular and a plural.

Choosing the level: ad group, campaign or account

Match the level to who is harmed if the negative is wrong. A mistake in an ad group costs a few clicks; a mistake in an account-level list can silence a whole campaign. Some illustrative placements for the same security client:

• "jobs", "salary", "careers": account-level list. No campaign wants a job seeker, and the words are unambiguous.
• "how to install": campaign level in the installation campaign. It is junk there, but a sibling campaign selling DIY kits would want it.
• "alarm" in the CCTV ad group, and "cctv" in the alarm ad group: ad group level, to steer each query to the ad group whose ads and landing page fit.
• A competitor's name that spends without converting: a brand exclusion at campaign level instead of a negative keyword.

Log every negative with its date, level, reason and approver. A flat log is enough, and it is the first thing a new team member will need when conversions dip.

Pages AI Max picks: URL and brand exclusions

Final URL expansion can send a query to any page it judges relevant, including pages you would never choose. In the first 28 days of the sample account, a look at the landing pages in the AI Max search term and ad combination view showed the careers page took 31 clicks and $52 with no conversions, because the page mentions installers. Blocking the queries behind those clicks would have needed dozens of negatives. Google's documentation recommends the opposite: when many queries bounce from one page, exclude the URL. Negative webpage criteria do that, and in standard Search ad groups they are campaign-level only.

Check the landing pages report, where a "Selected by" column shows the pages AI Max chose, and ask one question of each: would the client be happy for a first-time buyer to land here? If not, exclude it. Do the same for login pages, terms pages and old promotions. For competitors, a brand exclusion handles names that spend without converting, and keep a short negative list beside it for the misspellings and variants a brand list can miss.

The first clean-up in numbers, and what it is really worth

Illustrative figures for the same account. Last 28 days: spend $6,400, 3,900 clicks, 52 conversions, so $123 a conversion against a $120 target. The export holds 640 search terms with spend. Before any sorting, ask for the split by match source, because it shows how much of the money AI Max is steering:

A split like that does not prove AI Max is failing. AI Max can take credit for traffic the client's own keywords would have caught anyway, so its numbers are not all incremental, and a control is the only fair answer. It does tell the lead where to point the negative list: at the two AI Max sources. Now the AI sorts the 640 terms like this:

Of the $1,056 flagged as junk, $1,001 arrived through the two AI Max sources and $55 through the client's own keywords. The protect check pulls 9 terms costing $87 out of the junk pile. That leaves 129 terms costing $969, which is 15.1% of spend. They collapse into 34 negatives: 6 at ad group level, 19 at campaign level and 9 on the account list. The analyst holds 8 terms for a question to the client, and the lead approves the rest by class: job seekers, free and DIY, wrong product.

Now the sum the lead should do before any report goes out. If the campaign is limited by budget, the $969 does not disappear: it moves to other queries. At the campaign's $123 per conversion, that is at most 969 divided by 123, or about 7.9 extra conversions over 28 days, and probably fewer, because the next-best queries convert worse than the average. If the campaign is not limited by budget, the money simply goes unspent. Either way, call it spend redirected, never savings, unless the budget came down too. Check the result two weeks later with conversions per $1,000 of spend, which does not move when the budget does.

A weekly routine the lead can approve in ten minutes

After the first clean-up the weekly job is small. An illustration of the time budget for one account:

In a typical later week the illustrative sheet might show 190 terms reviewed, 17 proposals and one protected term removed. The lead approves 16 in ten minutes because the sheet shows classes and totals, not rows. Between reviews, an automated rule that emails you when a campaign's spend or cost per conversion jumps is often the first sign that AI Max has found a new pocket of queries; six rules that protect a small Google Ads budget lists the ones worth setting. Once a month, add three checks: read the account-level list against every campaign's products, look at the brand and URL exclusions, and compare the AI Max campaign with a control so the 14% claim gets tested on this client's data. The PPC audit template for new clients covers the equivalent first-time review.

The one-page sheet the lead should insist on

Anything longer than a page will not be read. This is a filled-in illustration of the first week's sheet:

The line that matters most is "converting terms blocked by the list". It should read zero every week, and a number above zero means the protect check failed or was skipped. The second most useful line is the spot-check result, because it tells the lead whether the AI is still classifying well.

Six questions to put to the team before the first approval

If you are the lead and have not looked inside an AI Max account recently, these six questions take ten minutes to ask and show how much control the account really has:

1. Which ad groups have search term matching switched off, and why those ones?
2. Which brand exclusions and URL exclusions exist, who added them and when?
3. How many account-level negatives are in use out of the 1,000 allowed, and which campaigns does each one touch?
4. What share of last month's spend went to keywordless matches, and what did it cost per conversion compared with the client's own keywords?
5. What is the control for the AI Max campaign, and how long will it run before anyone judges it?
6. When was the Converting sheet last refreshed, and by whom?

A team that answers all six from a spreadsheet, without opening the account, has a routine. A team that has to go and look is still setting up one.

When an account-level negative blocked a product line

An illustrative failure for the same client. The agency added "cheap" to the account-level list after it brought junk queries into the premium installation campaign. The client also ran a separate campaign for budget camera kits, and every query containing "cheap" stopped serving there. The first sign was not an alert. It was a week of lower enquiries for that product line, noticed when that campaign's conversions per $1,000 of spend dropped.

The fix had three parts. "Cheap" came off the account list and went onto the premium campaign as a campaign-level negative. The Monday checklist gained a step that reads every account-level addition against each campaign's products before it is approved. And the log gained a column for "campaigns affected", filled in at approval, not afterwards. A narrow negative that is slightly late costs a few clicks; a wide one that is wrong costs a product line.

Questions a PPC lead asks about AI Max negatives

Do negative keywords still work in AI Max campaigns?

Yes. Google's API documentation lists negative keywords at campaign and ad group level among the tools for irrelevant queries found in AI Max reporting, alongside brand exclusions and URL exclusions. What changes is how much traffic reaches you in the first place, because search term matching is on by default in every ad group, so the negative list does more of the controlling than it did before.

Can I stop AI Max from widening my exact and phrase keywords?

Yes. Google's help and API documentation both let you switch search term matching off for a single ad group, which stops the AI Max expansion there. It is a trade: you keep tight control in that ad group and give up the extra reach AI Max was bought for. Use it where a few precise terms carry most of a client's conversions.

Is it safe to paste client search terms into a chatbot?

Search terms can include names, addresses or phone numbers that people typed, so scrub them first and check your client agreement and the data controls on the plan you use. Send only the columns the task needs, and leave out account names. If the client's contract is silent, ask before you upload an export.

How much spend must a term burn before I add it as a negative?

My rule of thumb has two tiers. Clear junk such as job seekers or free-and-DIY queries can be blocked at any spend, using the narrowest wording that protects buyers. Borderline terms wait until they have spent about twice the target cost per conversion without converting, longer if the client's sales cycle is slow. Adjust it to the account's conversion lag.

How often should the negative lists be reviewed?

Weekly for the first month, then every two weeks if the junk share of spend stays low. Once a quarter, read every account-level negative against every campaign's products, because a term that was junk for one campaign can block another campaign's buyers.

Further reads

• AI Ad Copy Generator for PPC Agencies: Responsive Search Ads at Scale — Responsive search ads at scale, with the same review habit.
• Roofing Google Ads: Storm-Season Campaigns and Negative Keywords With AI — Negatives in a seasonal, high-intent trade.
• Performance Max Reporting: Explain Asset Results to Clients With AI — Turn a Performance Max asset export into a client explanation that says what the ratings mean, what they don't, and what to change next.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Google Ads API documentation for AI Max for Search campaigns and AI Max reporting, Google Ads Help on how AI Max works, on negative keywords and on account-level negative keywords, and Google's AI Max launch post (May 2025), as checked in November 2026.