How to Set Up AI Ad Governance Before Ads Go Live

The short answer: generation is no longer the slow part. Reading is. Once AI tools produce more ad variations than one person can read, governance is four checkpoints before launch, each with a named owner, plus a watch after launch. The four: which generators are switched on, the rules they must follow, a score on every ad against its page, and a human signature on every claim. AdAlign is the scorer for the third checkpoint. Its API on the Scale plan runs 1,500 evaluations a month (developers), and the free audit checks one pair at a time while you set the process up. The order: audit the generation settings, write the rules into the platforms, score every pair before it spends, have a person approve the claims, then monitor what goes live.

Disclosure: I earn a commission if you buy AdAlign through a link on this page, and this page recommends it. Every claim about AdAlign links to AdAlign's own pages.

The platforms write headlines from your landing page, Meta adjusts your images, and your own tools turn one brief into forty variants. Each output is a promise to a customer. Governance answers one question: who read it before it ran?

Size the problem first:

1. The reading load, worked out. For example, 40 variants a week is 40 x 52 / 12 = 173.3 a month. At 3 minutes to read each ad and open its page, 174 x 3 = 522 minutes. Nearly nine hours of reading a month, before any fix.
2. The controls you already pay for. Google's text guidelines let you add up to 25 term exclusions and 40 messaging restrictions per campaign for AI-written assets (text guidelines). Meta lists each Advantage+ creative enhancement, marks the ones that use generative AI, and lets you turn them off (About Advantage+ creative). Both are free.
3. The number that shows control. The count of AI-written assets you remove after launch. Google marks its generated assets as Google AI in the Added by column of the asset report (AI Max reporting). If removals fall month on month while volume holds, the checks before launch are doing their job.

Who checks what

1. Decide which generators are on

On Google, opting a Search campaign into AI Max turns on text customization and Final URL expansion by default. Together they let Google replace your final URL with another page, and write a headline to match it (Final URL expansion in Search). On Meta, Advantage+ creative can adjust uploaded media and text, and Meta says some enhancements may be on by default (About Advantage+ creative).

The generators do what they were built for. They feed the platform's auction more combinations to test. So the starting position is a set of variations your reviewer has not seen, and the review step has to be added from your side. Decide per campaign, record it in the launch ticket, and treat a change to it as a creative change.

2. Write the rules where the generator reads them

Google's text guidelines, in beta, take two kinds of rule. Term exclusions block exact words or phrases. Messaging restrictions describe concepts or a tone to avoid, and Google asks for them as direct requirements: a flat rule against implying a discount, not a wish (text guidelines). Two details catch teams out. Prices need the restriction "Don't generate assets with prices", because a term exclusion cannot block them. And new guidelines also remove existing generated assets that break them. Asset previews show up to 10 samples for a final URL before launch (text customization). A sample is not the full set, hence step 3.

Text guidelines can also be set through Google Ads Editor and the Google Ads API (text guidelines), so an agency can keep one rule file per client.

3. Score every pair before it spends

Here volume beats people, and AdAlign fits. An agent in your generation pipeline sends each new variant to POST /api/analyze-ad with the image URL, final URL and platform. It passes the copy as text in adTextFields, which AdAlign says scores message match better than reading copy off the image (analyze-ad reference). For an ad running across placements, such as Advantage+, the headline score is the weakest placement. One bad crop cannot hide behind a good average.

The routing rule the agent applies is a traffic light. Red, under 5, goes back to the generator with the element list. Orange, 5 to 6.5, goes to the media buyer. Green, 7 and above, moves on to the claims check. AdAlign counts spend at risk from scores below 5 and calls above 7 strong alignment (how it works). The human check: nothing reaches the platform on the score alone. A person sees every variant that passes, because the score measures fit with the page, not truth.

4. A person signs every claim

Prices, discounts, delivery promises, guarantees and regulated wording get a named approver, who checks them against the offer terms and the checkout. AdAlign applies the same principle to its own page rewrites: in Variants, price or discount claims ask for an explicit confirmation before they publish (how it works).

People in ads need their own line. Under Article 50(4) of the EU AI Act, deployers of an AI system that generates image, audio or video content constituting a deep fake must disclose that it was artificially generated or manipulated. A deep fake, there, is content that resembles existing persons, objects or places and would falsely appear authentic. The Regulation applies generally from 2 August 2026 (Regulation (EU) 2024/1689). Whether a synthetic presenter or product shot falls under it is for your counsel. Logging which assets are AI-generated is the part governance does now.

5. Watch what goes live

Pages change. A connected AdAlign account scores every new ad as it goes live, re-checks monitored pages daily and sends low-score and score-drop alerts to email or Slack (how it works). On Google, remove unwanted generated assets from the asset report. Bulk removal works through the API and Google Ads Editor, not the interface (AI Max reporting).

If procurement asks about AdAlign itself, its FAQ says it is actively working toward SOC 2 (FAQ, read 27 October 2026). Working toward is not certified. Write it that way in the vendor form.

Free first, then the scorer at volume

Steps 1, 2 and 4 cost nothing. For step 3 at low volume, the free audit scores one pair at a time, 5 a month on a free account (free audit FAQ).

Once variants arrive by the dozen, AdAlign, which pays me a commission, scores every live Meta and Google ad on every paid plan. On Scale it does the same through the API, from €1,290 a month on an annual contract, after a €1,500 paid 30-day pilot credited to the first invoices if you continue (pricing, checked on 27 October 2026). Try the first pair in the free audit, or have the pipeline step shown on your own account in the demo. The tool is summarised in the AdAlign entry. The single-launch version of the gate is in how to QA ad creative before launch.

FAQ

Do the platforms' previews replace a review?

No. The preview is a sample of up to 10 generated assets (text customization), while serving combinations are chosen per query, later. Tune the rules on the preview. Score what serves.

Who should own AI ad governance in a small team?

The paid media lead owns settings and scoring, and the offer owner signs the claims. Two people are enough, if whoever generated an ad never approves it alone.

Does AdAlign check whether an ad is labelled as AI-generated?

Its documented checks are the four alignment dimensions and their fixes (analyze-ad reference). Labelling is a separate checkbox in your launch ticket.