Addy
See what customers search forFind ad spend you're wastingRead paid and organic togetherPlan content from your search data
How it works
FoundersNew to search marketingProfessional servicesB2B SaaSDTC & e-commercePractitionersAgencies
PricingBlogToolsAbout
Sign InGet Started
  1. Home
  2. Blog
  3. Google Ads
  4. What to check before approving a Google Ads change an AI proposes
Google Ads

What to check before approving a Google Ads change an AI proposes

14 November 20269 min read
Hannah Reed

Hannah Reed

Digital strategist with over a decade in agencies and growth roles. Background in SEO and search strategy at EssenceMediaCom (WPP) and iCrossing (Hearst).

TL;DR

An AI-proposed change to a Google Ads account can be accurate in every line it shows and still cost money through a setting nobody chose. Three questions to answer from the account before approving: what was left to a default, what the rule does, and what the figures counted.

In this post

  1. How an AI assistant gets to change an account
  2. What the research says about reviewing AI suggestions
  3. What was left to a default?
  4. What does the rule do that the wording doesn't say?
  5. What did the figures count?
  6. A worked example
  7. After the approval
  8. What we still don't know
  9. Sources

An AI assistant that works on a Google Ads account usually stops before it writes anything. It shows the change, gives its reasons, and waits for a yes. The reasons tend to be tidy and the figures tend to be real, so the yes comes quickly.

A proposal can be accurate in every line it shows and still cost money, because the expensive part is often a setting nobody chose. Three questions catch a good share of it, and each can be answered from the account in a minute or two.

How an AI assistant gets to change an account

There are three common routes into an account, and the approval step is different on each.

The first is Google's own connection between AI assistants and Google Ads. Its developer guide, last updated on 13 November 2026, describes the mode as "Read and partial mutate". The writes it supports are creating assets, uploading videos and managing labels, and the guide says: "It also cannot create, modify, pause, or enable campaigns, ad groups, ads, budgets, or bids." Each write runs in two steps: a dry run that returns "human-readable change_items entries" to read, then a commit that goes through only if the request matches what was shown.

The second is a general assistant with a connection added. Anthropic's help centre says connectors let Claude "access your apps and services, retrieve your data, and take actions within connected services", and that connecting one grants permission "to access and potentially modify data within that service based on your account permissions". Tool calls come with approval requests, and the guidance for custom connectors is to click "Allow always" only for "a server and tool that you trust to run unsupervised". On Team and Enterprise plans an owner can set each tool to "Always allow, Needs approval, or Blocked". One mode behaves differently: during Research, Claude "can invoke tools from your connectors automatically without further approval", and the same page advises switching off any tool that can write.

ChatGPT works along similar lines. OpenAI's guide to custom connections says "Write actions by default require confirmation" and asks the user to "Carefully review the tool input which will be sent to a write action". A tool counts as read-only only when its developer has marked it that way: "Tools without this hint are treated as write actions." An approval can also be remembered for the rest of a conversation, which OpenAI suggests keeping for tools the user trusts.

The third route is a tool built on a model's API, which covers a lot of specialist ad software and anything a team writes in-house. Here the model never touches the account. Anthropic's documentation says that for tools a developer defines, "your application runs each call". OpenAI's function calling guide says a function definition lets the model "pass data to your application, where your code can access data or take actions suggested by the model". On this route an approval step exists only where someone built one, and what it shows is the builder's choice. OpenAI's hosted option for remote servers is a partial exception: it "defaults to requiring approvals of each MCP tool call", and a developer can turn that off.

Whether any of these reaches Google Ads depends on what's connected: Google's server, a third-party one, or code written for the purpose. The protocol most of them share asks that "there SHOULD always be a human in the loop with the ability to deny tool invocations". Across all three routes, then, the weight sits on the moment of approval, and an approval can only show what the request contains.

What the research says about reviewing AI suggestions

The natural way to review a proposal is to read the explanation and judge whether it holds together. The research on people working with AI suggestions gives a reason to be careful with that habit, and it doesn't all point one way. Bansal and colleagues (CHI 2021) found that explanations "increased the chance that humans will accept the AI's recommendation, regardless of its correctness". Vasconcelos and colleagues (CSCW 2023) found that explanations can reduce over-reliance, and across five studies with 731 participants it fell when checking the AI's answer was made cheaper.

These were lab tasks such as mazes and text classification. None involved an ad account, and the size of the effect in paid search is unknown. The practical reading is modest: a persuasive explanation says little about whether the change is right, and a check gets done when it's quick. So the three questions below leave the explanation alone and go to things the account can answer.

What was left to a default?

Every setting a proposal doesn't mention still gets a value. Some are filled in by Google and some by the tool, and the proposal is usually silent about both.

Campaign status is the plainest case. Google's API reference says of a campaign's status: "When a new campaign is added, the status defaults to ENABLED." Google's own code sample for creating a campaign carries a comment advising the opposite: "Set the campaign to PAUSED when creating it to prevent the ads from immediately serving." A proposal for a new campaign that doesn't state a status could go either way, depending on how the tool was built.

Location works the same way. Google's location targeting guide labels the wider option, which reaches people in the area and people who have shown interest in it, as the "Recommended default" and notes that it "provides the broadest reach". For a business that only serves customers who are physically nearby, the narrower option is the one that matches the brief, and a proposal that says "targeting Leeds" doesn't say which was used.

Match type, the network settings and the bid on a new ad group belong on the same list. The question for each is the same: where the proposal is silent, what value will the account end up with?

What does the rule do that the wording doesn't say?

Some settings behave differently from their plain reading, and a proposal written in plain language inherits the gap.

A daily budget is an average. Google's help page says a campaign "might spend up to twice your average daily budget" on a given day, and that over a month it will spend "no more than 30.4 times" that figure. A proposal to set £40 a day is a proposal to allow £80 on a busy day and £1,216 over the month. Since June 2026 the pacing behind that monthly figure has also changed for scheduled campaigns.

Negative keywords are narrower than they read. Google's page says "Negative keywords won't match to close variants or other expansions", and gives the example that excluding "flowers" still allows an ad on a search for "red flower". A proposal to "exclude job seekers" by adding "jobs" leaves "job" open. The list in the proposal is the whole exclusion, with no margin around it.

A maximum CPC is a ceiling on what a click can cost. The figure to set it against is what a click is worth to the business.

What did the figures count?

Proposals quote the account's numbers, and the numbers are usually right. What varies is which column they came from and which days.

Google Ads keeps two conversion counts. The "Conversions" column holds primary actions only; a secondary action "is excluded from the 'Conversions' column" and appears under "All conversions". A cost per conversion worked out from the wider column can look far better than the one bidding runs on. The same care applies to what each conversion action records in the first place.

Recent days are the other gap. Clicks and cost refresh within hours, while Google says freshness for conversions "varies significantly depending on your attribution model and conversion type". A proposal that reads the last seven days can be comparing complete spend with conversions that are still arriving. Opening the same date range and the same column in the account settles both points.

A worked example

The figures here are invented to show the three questions at work.

An assistant proposes a new Search campaign for a regional service business: £40 a day, twelve phrase match keywords, the negatives "free" and "jobs", and a rationale that the existing campaign produced 38 conversions in the last 30 days at £21 each.

  • Left to a default: the proposal states no status and no location option. Unless the tool sets them, the campaign starts live and reaches people who are interested in the region as well as people in it.
  • The rule behind the wording: £40 a day allows £80 on a single day and £1,216 across the month. The negative "jobs" doesn't cover "job".
  • What the figures counted: the 38 came from "All conversions". Fourteen were brochure downloads set as a secondary action, so the "Conversions" column shows 24. The same £798 of spend then works out at £33.25 a conversion.

Nothing in the proposal was false. Each of the three answers changes whether it gets approved as written.

After the approval

The account keeps its own record. Change history "lists the changes made to your account, campaigns and ad groups during the past two years", and Google notes that changes made through the API "might show 'Google Ads API' or a specific tool name in the user column". That makes it possible to see afterwards what was changed, by which tool, and what the value was before. It's also the place to confirm that what was applied matches what was approved.

What we still don't know

How often AI-proposed changes to ad accounts are wrong, and how often reviewers catch them, isn't measured by anything cited here. The research comes from other tasks. Google's guide states what its connection can't change and doesn't give the reasoning. And the three questions are a floor. They catch settings that were never chosen, and they have nothing to say about whether the strategy behind a proposal is a good one.

A long proposal raises a separate problem. Fifty bid changes can't each get three questions, and nothing cited here says which rows deserve the attention. Reading the largest changes and the ones that look unlike the rest is a common practice. Whether it's enough is an open question.

Sources

  • Google Ads API documentation, Google Ads MCP server: developer integration guide, last updated 13 November 2026.
  • Google Ads API reference, Campaign, and the Add campaigns code sample.
  • Google Ads API documentation, Location targeting.
  • Google Ads Help: About overdelivery and your average daily budget; About negative keywords; About conversion goals; About data freshness; About change history. Read on 14 November 2026.
  • Anthropic, Claude Help Center: Use connectors to extend Claude's capabilities; Get started with custom connectors using remote MCP. Claude Platform Docs: Tool use with Claude. Read on 14 November 2026.
  • OpenAI developer documentation: Add custom MCP server; Function calling; MCP servers. Read on 14 November 2026.
  • Model Context Protocol, specification 2025-06-18, Tools.
  • Gagan Bansal and colleagues, "Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance", CHI 2021.
  • Helena Vasconcelos and colleagues, "Explanations Can Reduce Overreliance on AI Systems During Decision-Making", CSCW 2023.
Previous

How to set a manual CPC bid: start from what a click is worth

Next

How to check a Google Ads recommendation before you apply it

Related articles

Google Ads

How to review a negative keyword list in Google Ads

A negative keyword list can let through searches it was meant to stop, and stop searches the business wants. How negatives match, which entries to check first, and why adding every variant in advance isn't the answer.

16 November 2026·8 min read
Google Ads

How long to wait before judging a change in Google Ads

Google Ads reports a conversion on the day of the click, so a recent week shows all of its cost and only some of its conversions. How to measure the delay in your own account with the Days to conversion segment, and when a week is old enough to compare.

15 November 2026·4 min read
Google Ads

Should you bid on your own brand name? The evidence, and a way to check

Experiments on brand ads range from no measurable benefit to a strongly positive return, and the one study that separated the cases found competitors on the page made the difference. What to look at first, and what a pause test can and can't show.

15 November 2026·7 min read

Tools

  • SEO checklist
  • Technical SEO crawler
  • Search benchmarks
  • Glossary
  • All tools

Guides

  • Google Ads for small business
  • Create a Google Ads account
  • Verify Search Console
  • All guides

Topics

  • Google Ads
  • Technical SEO
  • Search strategy
  • Conversion tracking

Product

  • How it works
  • Pricing
  • About
  • Get started

Solutions

  • See what customers search for
  • Find ad spend you're wasting
  • Read paid and organic together
  • Plan content from your search data

Who it's for

  • New to search
  • Professional services
  • B2B SaaS
  • E-commerce

© 2026 Addy

·

Not affiliated with Google

PartnersContactPrivacyTerms