Estimate how much delivery disappears before adding exclusions: the impact of P-MAX negative keywords and Meta Ads placement exclusions

Translated from the Japanese original on mureo.jp.

When you find a search term or placement that performs badly, the first thing you probably think of is adding an exclusion. In campaigns like P-MAX and Advantage+, where choosing where ads run is left to the platform’s automation, exclusions are one of the few ways the operator can impose limits from their side.

An exclusion, though, is an operation that reduces delivery. How much it reduces can be calculated from recent results before you add it. Add it without calculating, notice only after delivery breaks down, and removing the exclusion may not bring the original delivery back.

Here we set out three records of what happened after exclusions were added, run the same estimate on our own Meta Ads results by placement, and then write up how exclusions work on Google Ads and Meta Ads, along with steps you can take before adding them with read-only access. How to connect Claude Code to each platform is covered in Connect Google Ads to Claude Code and Connect Meta Ads to Claude Code.

What an exclusion reduces

When you add an exclusion, the impressions that were going to that term or placement go away. With the impressions gone, the clicks, spend and conversions that came from them go away too. So if you retrieve how many impressions, clicks, spend and conversions an exclusion candidate had over the last 30 days, you get a rough figure for how much adding the exclusion would reduce.

It stays a rough figure because the freed-up budget may be used on the remaining destinations. In an automated campaign, the budget for what you excluded does not necessarily sit unused; it may move to the terms and placements that remain. There is no guarantee that results where it goes will be better than where you excluded.

There is one more reason to look at the reduction first: removing an exclusion does not necessarily restore things. The second record in the next section is an example.

What happened after exclusions were added

Here are three first-person posts we came across on X. In each case, we have not verified that the account is real or what the business is.

About 100,000 yen spent on P-MAX with 0 conversions
A post from July 2026 by someone running their own service. They spent about 100,000 yen on P-MAX and got 0 conversions. Looking at the search terms, delivery had drifted toward terms unrelated to the service, and they wrote that adding exclusions could not keep up.

After adding a large number of exclusions, the budget flowed to display
An operations log from September 2026. Judging that P-MAX’s expansion was too broad, they added a large number of exclusions, and the budget flowed to display inventory that brought few conversions. Even after removing the exclusions and switching the light conversions to secondary, delivery did not come back.

The surface nearly stopped on single-month CPA correlated with sales
A post from September 2026 by an executive at a company with its own media and social ads businesses. They matched the daily cost by surface of a P-MAX campaign for ecommerce against sales on an online marketplace. Looking only at the cost per acquisition (CPA) in the platform UI for September alone, the call would be to stop the YouTube surface, but looking at 235 days with seasonal and promotion days removed, the assessment reversed, and YouTube was the only surface positively correlated with sales.

What happened in the three records differs. The first is a case where exclusions could not keep up, the second a case where adding exclusions moved the budget to another surface and removing them did not bring it back, and the third a case where a surface nearly excluded was in fact driving sales. In the second and third, laying out what an exclusion would remove before adding it might have changed what the decision was based on.

Our own Meta Ads results by placement

From July 27 to August 31, 2026, we ran Meta Ads to compare messaging on mureo’s landing page. Total spend was 113,587 yen. Results by placement (publisher_platform) were as follows.

Placement	Impressions	Clicks	Spend	CTR	Conversions	CPA
Instagram	7,832	179	61,607 yen	2.29%	2	30,804 yen
Facebook	4,069	146	45,380 yen	3.59%	1	45,380 yen
Audience Network	2,968	262	6,600 yen	8.83%	1	6,600 yen

The totals are 14,869 impressions, 587 clicks and 4 conversions. Conversions are as measured by the Meta pixel.

Apply the rule “drop the placement with the worst CPA” to this table, and the one to drop is Facebook. Our tool’s placement analysis also returned that Facebook’s CPA was 6.9 times Audience Network’s, so excluding it should be considered. But dropping Facebook would remove about 40% of spend (45,380 yen of 113,587 yen) and 1 of the 4 conversions.

On the other hand, apply the commonly cited rule “drop Audience Network because its click quality is low,” and the one to drop is Audience Network. Spend falls by only 5.8%, but 44.6% of clicks, 20.0% of impressions and 25% of conversions (1 of 4) disappear. Audience Network’s CTR of 8.83% is 2.5 to 3.9 times that of the other placements, which is grounds to suspect accidental taps. Even so, a high CTR alone does not establish that those clicks have no value.

For the two candidates, the share that disappears if you drop each, across four metrics, is as follows.

Placement dropped	Impressions	Clicks	Spend	Conversions
Facebook	27.4%	24.9%	40.0%	1 of 4
Audience Network	20.0%	44.6%	5.8%	1 of 4

From the same results, depending on which rule you apply, you reach the conclusion to drop opposite placements. On top of that, conversions total 4, not a count that can statistically support either decision. Since these results cannot tell you which rule is right, it is safer to first lay out “what disappears, and how much, if you drop it” across impressions, clicks, spend and conversions, and decide after looking at that.

From where we sit building ad-ops tooling, we have built a feature into mureo that produces this estimate, and the numbers above came from it, analysis_exclusion_impact_preview. Before you add an exclusion, it returns what share of your own account’s recent delivery would disappear, both for the current candidate and as a cumulative figure combined with existing exclusions. When there is not enough data it returns “unknown,” and does not treat that as “no impact.” If you write a ceiling into the guardrails in your operating strategy file (STRATEGY.md), exclusions that exceed that ceiling are stopped before they are added.

How far P-MAX negative keywords reach

In Google Ads P-MAX, the limit on campaign-level negative keywords was raised from 100 to 10,000, and they became available in all P-MAX campaigns (Google Ads Help). Campaign-level negative keyword lists can also be applied to P-MAX. The limit on account-level negative keywords is 1,000 (Google Ads Help).

The higher limit makes it easier to add large numbers of exclusions while reviewing search terms. However, P-MAX negative keywords apply only to Search and Shopping inventory, not to YouTube, Display, Gmail or Discover. Negative keywords also do not match close variants, so if you want to stop similar terms too, you add those terms separately.

As a result, excluding a large number of search terms narrows only the Search and Shopping side, and the surfaces negative keywords do not reach stay as they are. If the campaign budget does not change, the freed-up amount may flow to the remaining surfaces (such as Display). The second record in the previous section, where the budget flowed to display, is consistent with this mechanism. That said, there is no way to confirm from outside what actually caused it in that record, so we cannot go as far as cause and effect.

Before adding negative keywords to P-MAX, look not only at the candidate terms’ results but also at the share of the whole campaign’s cost taken by Search and Shopping. That tells you how large the range the exclusions narrow is, and how large the range they do not.

Meta Ads placement exclusions, and delivery to excluded placements

In Meta Ads, besides setting placement exclusions per ad set, there is a setting to remove placements such as Audience Network at the account level.

In October 2025, a mechanism called “limited spend” was introduced in the Marketing API. It allows delivery to excluded placements of up to 5% of spend (Meta for Developers blog). In the API it is not turned on automatically and works only when explicitly specified. We have not been able to confirm what the default behavior is for ad sets created in the Ads Manager UI. If you see a small amount of spend on a placement you thought you had excluded, check whether this setting is on.

There are also reports that a change making placements impossible to choose manually, depending on the campaign objective, is being rolled out in stages. We have not confirmed this in our own account, so before dropping a placement, you need to check in the platform UI whether placements can still be chosen in your own campaign.

Steps before adding an exclusion

All of the following steps can be carried out with read-only access.

Get the last 30 days of results for each candidate: for each exclusion candidate (term or placement), retrieve impressions, clicks, spend and conversions for the last 30 days. If an AI agent does the retrieval, have it add to its answer which account and which period it looked at. Mix-ups over where it is looking and the period are covered in Can you trust the ad numbers an AI agent gives you?.
Work out the share that disappears across four metrics: work out what percentage of the whole campaign disappears if you drop the candidate, for each of impressions, clicks, spend and conversions. Look at only one metric and, as in the Facebook and Audience Network example above, you miss that you are cutting deeply on another.
Work it out as a cumulative figure with existing exclusions too: look at the cumulative figure, adding what existing exclusions already narrow to the current candidate’s share. Exclusions that are each small can, stacked up, narrow the range of delivery a great deal.
If conversions are too few, hold off on the decision: when there are only a handful in total, as in our example, extend the period and wait for the count to grow, or test small in a single ad set or campaign.
Record the date and reason, and decide in advance the criteria for reverting: record in your change log the date you added the exclusion, the reason, and the values of the four metrics at that time. At the same time, decide before adding it which numbers, moving how, would make you revert. The platform history does not keep the reason for a change; how to build a record you keep yourself is in Keeping your own change log for ad accounts.

As for the reverting criteria in step 5, since removing an exclusion may not restore things, as in the second record, it is safer not to put this off thinking “I can just revert.”

If you let an AI agent add exclusions, decide separately from approval how many it may add at once and a ceiling on the share it may narrow. Relying on approval alone makes it easy to pass a lined-up list of exclusion candidates all at once. How to place ceilings and approval is covered in Is it safe to hand ad ops to Claude Code?.

If you have put the exclusion candidates and their recent results into a table but cannot decide which to drop, you can consult us with that table as is through the contact form. With the list of candidates and the exclusions already in place, it is easier to get into specifics.

Frequently asked questions

How many negative keywords can I add?

The limit is 10,000 at the P-MAX campaign level and 1,000 at the account level. But that is the limit on how many you can add, not a guide to how many you should. Rather than the number, it is more reliable to judge by how much of the campaign’s impressions and clicks the cumulative exclusions you have added are narrowing.

I added negative keywords to P-MAX and display costs went up. Why?

P-MAX negative keywords apply only to Search and Shopping inventory, not to Display, YouTube, Gmail or Discover. As the Search and Shopping side was narrowed, the budget may have flowed to the remaining surfaces. Since the cause cannot be pinned to one thing, line up the cost by surface before and after the day you added the exclusions and check.

Should I drop Audience Network?

General rules alone cannot decide it. In our Meta Ads, Audience Network was 5.8% of spend but accounted for 44.6% of clicks and 1 of 4 conversions. A CTR far higher than other placements is grounds to suspect accidental taps, but before dropping it, work out what disappears across four metrics and check whether the conversion count is enough to decide on.

If I remove an exclusion, does delivery go back to how it was?

Not necessarily. In the record of an operator who added a large number of exclusions to P-MAX and saw the budget flow to display, delivery did not come back even after removing the exclusions and switching the light conversions to secondary. Do not assume you can revert; estimate what disappears before adding, and decide the criteria for reverting.

How should I decide when conversions are few?

Hold off on the decision and extend the period, or test small in only some ad sets or campaigns. In our example, conversions totaled 4, and whether you looked at CPA or CTR reversed which placement to drop. While the count is small, it is sensible to look at the share that disappears in impressions, clicks and spend, which have larger counts, and stay within a range that does not cut delivery heavily.

Summary

An exclusion is an operation that reduces delivery, and how much it reduces can be calculated from recent results before you add it. In our Meta Ads, judging by CPA led to dropping Facebook and judging by click quality led to dropping Audience Network, and dropping either would remove 1 of 4 conversions. P-MAX negative keywords apply only to Search and Shopping, and Meta Ads has a setting that allows limited delivery to excluded placements.

Before adding an exclusion, lay out how much of impressions, clicks, spend and conversions disappears, both per candidate and as a cumulative figure with existing exclusions; hold off if the count is not enough; and record the reason you added it and the criteria for reverting. That makes it easier to avoid noticing only after delivery has broken down.