Google Ads sees the spend. Google Analytics sees the sessions and conversions. Ask Advisor can now bring those signals into one conversation, spot a shift, build a visual report, benchmark performance, and recommend a move.

That sounds useful. It also makes a bad marketing habit much easier: treating a confident explanation as a checked conclusion.

The fix is not a ban on the tool. Give it a lane, require evidence, and keep a person on the last click.

Let the AI shorten the investigation. Do not let it skip the investigation.

What changed in Google Ads and Analytics

On August 10, Google announced new Ask Advisor capabilities for English-language accounts in beta. Google Analytics can surface AI-generated homepage summaries and carry a selected data card into Ask Advisor. Google Ads can create custom insights from a prompt. Prompt-built dashboards are available in Ads and are coming to Analytics, while Analytics is adding comparisons against anonymized averages from similar businesses.

Google’s earlier Ask Advisor introduction describes a broader cross-product agent that can pull information from Merchant Center, set up a campaign in Google Ads, combine Ads and Analytics data, explain results, and recommend the next action.

The important change is the distance between a question and a campaign decision. It just got shorter. Useful, yes. Short distances also make it easier to step over missing context.

First, put requests into three lanes

Before anyone starts prompting, decide what Ask Advisor is allowed to do. A simple three-lane rule works for a small team:

  • Read: summarize a performance shift, find a segment, or build a report. A marketer can use this lane freely, then verify the important numbers.
  • Draft: propose a budget change, audience, campaign structure, or test. The output goes into a brief. Nothing changes in the account.
  • Act: create or modify a live campaign, goal, budget, bid, audience, or asset. This requires named human approval and a change record.

Do not call all three “using AI.” Reading an anomaly and raising a daily budget are different risk classes. Permissions and review should reflect that.

Ask a decision question, not a fortune-cookie question

“How can we improve performance?” will invite a broad answer assembled from whatever the system considers important. Give the tool a real decision instead:

Better prompt: “For Wisconsin campaigns from July 1 through August 15, compare qualified lead rate and cost per qualified lead with the prior six-week period. Separate brand and nonbrand traffic. Exclude internal form tests. Identify the three largest changes, show the source metrics, and propose one reversible test. Do not make account changes.”

That prompt supplies the market, dates, outcome, comparison, exclusions, output, and permission boundary. The AI still may miss something. At least you gave it a defined job.

Make every recommendation carry a receipt

Ask Advisor can explain the “why” behind performance, according to Google. Treat that explanation as a hypothesis until the underlying report supports it.

Require every meaningful output to include:

  • the exact account, property, campaign, channel, audience, and geography in scope;
  • the current and comparison date ranges;
  • the metric definition, including which conversion actions count;
  • the absolute numbers as well as percentage changes;
  • known exclusions, tracking changes, and missing data;
  • the native report or view where a person can reproduce the result;
  • a confidence label: observed, inferred, or speculative.

This is especially important when benchmarks enter the conversation. An anonymized peer average can provide direction, but it cannot know your margin, sales capacity, lead quality, seasonality, or which firms Google placed in the comparison set. A benchmark is a prompt for another question. It is not a target handed down on a stone tablet.

Reproduce the finding before debating the advice

Open the native Ads or Analytics report and rebuild the important comparison. Check filters, attribution settings, conversion definitions, consent gaps, campaign changes, and the dates of promotions or outages.

A 30 percent improvement can mean 10 leads became 13. A falling cost per lead can hide a collapse in qualified leads. A traffic spike may be a tagged email send, referral spam, or one enthusiastic employee refreshing the new landing page.

SigServe’s marketing causality check is useful here. It separates what happened from what the dashboard wants credit for. If the tracking foundation is shaky, use the pixel and privacy audit before asking an agent to make increasingly precise conclusions from increasingly questionable inputs.

Turn the recommendation into a small, reversible test

Once the evidence holds, shrink the proposed action. Define:

  • Hypothesis: what should change, for whom, and why.
  • Primary measure: the business outcome that decides the test.
  • Guardrails: spend, lead quality, frequency, geography, brand terms, and any non-negotiable exclusions.
  • Duration: enough time or volume to learn without letting a weak test wander through the quarter.
  • Stop rule: the condition that ends the test early.
  • Owner: one person responsible for approval and review.

“Increase budget 20 percent” is an action. “Increase the nonbrand campaign budget by 10 percent for 14 days while qualified cost per lead stays below $180 and impression share lost to budget remains above 15 percent” is a controlled test.

Keep a change log boring enough to trust

For every action, record the Ask Advisor prompt, its recommendation, the supporting report, the human approver, what changed, when it changed, and the review date. Add campaign annotations where the platform supports them. Use consistent UTM naming when the test sends traffic outside the account.

This is not paperwork for its own sake. Two weeks later, the team needs to know whether performance moved because of the approved test, a tracking edit, a new offer, a sales follow-up problem, or three unrelated changes made on a Friday afternoon.

What to ignore

Ignore the pressure to use every new capability immediately. Ignore a pretty auto-generated chart until the definitions check out. Ignore a peer benchmark that cannot be tied to your economics. Ignore recommendations that solve for platform activity while the business needs qualified revenue.

Also ignore the fantasy that one conversational interface eliminates marketing expertise. The interface can reduce report-building time. It cannot decide which tradeoff the business should accept, notice every broken handoff, or own the result.

The ten-minute approval check

  1. Is the question tied to a real business decision?
  2. Can we reproduce the finding in a native report?
  3. Are dates, segments, definitions, and exclusions visible?
  4. Did we check business context the platform cannot see?
  5. Is the proposed action reversible, bounded, and measurable?
  6. Does one named person approve the change and own the review?

If any answer is no, keep the recommendation in the draft lane. Ask Advisor can wait. Your budget has already demonstrated an impressive ability to leave without supervision.

Is the AI speeding up decisions or just speeding past the checks?

SigServe can build the reporting rules, approval lanes, and test plans that keep useful automation attached to business judgment.

Build the Workflow →