Google’s AI Search toggle needs a test plan, not a gut decision
By Greg Nowak. Last updated 2026-07-11.
Google’s new generative AI control in Search Console looks like a simple switch: let your content appear in covered AI features, or exclude it. For a business that relies on search discovery, the decision is much bigger than the interface suggests.
Leave inclusion enabled and your content remains eligible to appear as links and help ground responses in AI Overviews, AI Mode and generative AI features in Discover. Exclude it and you give up that eligibility, together with any impressions and traffic those features might generate. Publishers and content-led businesses are weighing reach against control and the value they receive from being visible.
That is not a decision to make site-wide on instinct. Treat it as a measured test: capture the baseline, define what a useful outcome looks like and allow enough time to separate a real effect from normal fluctuation.
Be precise about what the control changes
Google says the control is being tested with a subset of website owners. Where it is available, inclusion is the default. Excluding a site prevents its links and content from appearing in the covered generative AI features or being used as input for responses and previews in those features.
The scope is deliberately narrow. Google says the setting is not a ranking or inclusion signal for other parts of Search. It does not control AI model training, either, and it does not replace separate choices for services such as Google Ads or Merchant Center. A governance record should describe the decision accurately. Saying “we opted out of Google AI” would claim far more than the setting actually does.
Search Console’s property structure may offer room for a smaller test. Google documents inheritance between parent and child properties, while allowing owners to override the inherited choice for an eligible child property. If commercially distinct content already sits in a suitable verified property, you may be able to test that content without changing the whole site. Check the property hierarchy and owner permissions first; a folder is not automatically an independently controllable property.
Do not treat the moment someone clicks the toggle as a clean dividing line in the data. Google says exclusion generally takes a few days. Content should be excluded within one or two days after the control goes live, but caching and propagation may delay some removals. Mark that transition period in the analysis instead of assigning it to a neat “before” or “after” bucket.
The report shows exposure, not business value
The Generative AI performance report is the obvious place to establish a baseline. It reports impressions from AI Overviews and AI Mode, grouped by page, country, date and device. The data can be exported, so the team can preserve a proper snapshot instead of relying on screenshots or recollection.
An impression only tells you that a link to the site appeared in a supported generative feature. It does not tell you whether anyone visited, enquired, subscribed or bought something. The report also excludes Search Labs experiments, retains familiar Search Console constraints such as the 1,000-row limit, and may show different chart and table totals because the two views aggregate data differently.
Manual spot checks are not a reliable substitute. Research into generative-search measurement found that the same queries, submitted at different times, could produce different responses and cite different sources. The experiments examined Perplexity Search, OpenAI SearchGPT and Google Gemini, not this specific Search Console report, so they are not a direct audit of Google’s new control. They do show why a few casually repeated prompts are weak evidence. AI visibility needs structured, repeated observation.
Build a test around the business decision
Start with the question the test must answer. One useful version is: does inclusion create commercially valuable discovery for our important content without giving up more control than we are prepared to accept? That gives the team something concrete to assess. “Let’s see what happens” does not.
Before changing anything, export the available generative AI report. Preserve page-level data as well as country, device and date views. Record the export date, reporting period, Search Console property, current setting and whether the setting is inherited. Save ordinary Search performance and relevant analytics for the same period. AI impressions make little sense in isolation.
Then classify the pages by what they do for the business. A service page, product comparison, original research article and general explainer may all earn impressions, but those impressions do not carry equal value or risk. Note which pages support leads, subscriptions, direct revenue, brand authority or customer support. Flag content whose economics depend on controlled access, licensing or clear attribution.
| Stage | What to record | Why it matters |
|---|---|---|
| Baseline | AI impressions by page, country, device and date; ordinary Search performance; referrals and conversions | Separates the effect of the setting from an existing trend |
| Page classification | Commercial role, originality, access model and governance sensitivity | Prevents low-value and high-value impressions from being treated alike |
| Change log | Property, inherited or overridden setting, timestamp, owner and rationale | Confirms what changed, where and under whose authority |
| Transition window | Expected delay, caching notes and observed propagation | Stops the team from judging the result before the change has taken effect |
| Monitoring | Repeated exports, referrals, leads and ordinary Search trends | Connects visibility with actual business behaviour |
| Review | Pre-agreed success, stop and reconsideration criteria | Keeps the final decision tied to evidence rather than preference |
If the test involves changing the setting, annotate the change and account for Google’s implementation delay. Where possible, avoid redesigning templates, altering internal links, rewriting major content sections or changing conversion tracking during the same period. If other work cannot wait, log it. Otherwise, you may end up attributing its effect to the AI control.
Review the data on a regular cadence, not after one unusual day. Look at page groups as well as the site-wide total. A steady headline number can hide declining visibility for commercially important pages while low-value content grows. Track ordinary organic performance alongside direct and referred visits, enquiries, sign-ups or the conversion events the organisation already trusts. Search Console supplies the exposure measure; your own analytics supplies the commercial context.
Agree on the decision rules in advance
Set the criteria for inclusion, exclusion or further testing before the results arrive. Continued inclusion may make sense when valuable pages receive sustained visibility and that exposure accompanies useful visits or outcomes. Exclusion may deserve serious consideration when content policy, licensing terms or negotiating strategy matter more than the discovery benefit you can measure.
Some tests will be inconclusive. That is a reason to extend the measurement period or narrow the scope, not to manufacture certainty from thin data.
The regulatory context makes clear that this belongs with governance, not just technical SEO. The UK Competition and Markets Authority presented the control as a way to give publishers more meaningful choice and stronger bargaining power over content used in AI Search features. It does not tell any individual website which setting to choose. It does mean the choice needs an accountable owner, a recorded rationale and evidence that matches the commercial stakes.
Google has supplied a control and an early visibility report. The business case still has to be built. A capable freelance operator can bring the technical configuration, content priorities, analytics and policy considerations into one manageable experiment, then help the team revisit the choice as the rollout and reporting develop.
Related on GrN.dk
- Support bots need a deletion test before they learn the old help center
- When Google can call the business, your local data stops being cosmetic
- WordPress 7.1's Media Pipeline Needs a Real Plugin and CDN Test Pass
Need help with this kind of work?
Plan your AI Search visibility test Get in touch with Greg.