Google has finally put a number beside your visibility in its AI search products. It has also added a switch that can make that visibility disappear. I would measure the traffic before touching it.
What changed in Search Console
Google says its new generative AI controls and reporting completed a worldwide rollout on August 31, 2026. Eligible Search Console properties can now see impressions from AI Overviews, AI Mode, and AI Overviews in Discover. Reports can also break those impressions down by page and country.
The same property-level area includes a control that can exclude site links and content from those generative search experiences. Google is explicit about the tradeoff: exclusion also removes the traffic and impressions that those features could send.
This is separate from Google-Extended. Google-Extended controls whether content may help improve certain generative AI models. The new Search Console switch controls whether content appears in Google’s generative search experiences. Mixing those two settings could produce a decision you did not intend.
The report answers one question and leaves out another
The new data answers: where is my site being shown? It does not yet answer the more useful commercial question: what did those impressions do?
| Decision input | Available now | What is still missing |
|---|---|---|
| Visibility | Impressions across supported AI experiences | Position and answer-level prominence |
| Content | Pages receiving impressions | The exact passage or citation used |
| Geography | Country breakdown | User intent and query detail |
| Business outcome | Not enough by itself | Clicks, conversions, assisted revenue, and lead quality |
That missing click layer matters. A page with 50,000 AI impressions and no measured visits is a different asset from a page with 2,000 impressions that sends qualified buyers. Impression volume alone is not a reason to stay in or opt out.
Price the opt-out before you use it
I would treat the switch as a business experiment, not a privacy reflex or an SEO tactic. Google says the setting is not a ranking signal for non-generative search, but removing content from AI experiences still changes where people can discover you.
Start with a simple exposure calculation:
Estimated monthly value at risk = AI impressions × estimated click rate × conversion rate × value per conversion.
The weakest number is the click rate because Google does not provide it in this report. Use a low, base, and high assumption instead of pretending one estimate is exact. For 100,000 monthly impressions, a 0.1% click rate means 100 visits. A 1% rate means 1,000. If 2% of those visitors convert at $80 each, the range is $160 to $1,600 per month before assisted conversions.
That calculation does not prove the traffic exists. It shows how much uncertainty you are accepting when you flip the switch without an observation period.
A four-week test for publishers
- Export the generative AI report each week. Record total impressions, top pages, and countries.
- Tag the pages by commercial role: news, guide, comparison, tool page, newsletter acquisition, or direct sale.
- Compare referral changes in analytics with branded search, direct visits, conversions, and assisted conversions.
- Identify sensitive or low-value content that may deserve exclusion without removing the whole property.
- Recalculate the value-at-risk range before changing the property-level control.
If the report is not visible yet, check whether the property has enough qualifying impressions and whether you are using the correct property. A rollout can be global while an individual report remains empty.
The early research makes the choice harder
A preregistered field experiment involving 1,100 participants reported that removing AI search features increased publisher clicks. The authors also found that an AI Mode-only condition reduced click-through rate and some measures of user experience and trust. It is a preprint, not peer-reviewed consensus, and it does not reproduce every real Google query environment.
Still, the direction is worth taking seriously. A search answer can satisfy a user while giving the source little or no traffic. That is why the new visibility report needs to be paired with business outcomes, not celebrated as a complete analytics product.
When exclusion can be the right decision
Some pages should not be optimized for maximum exposure. A paid research product, licensed database, private community resource, or page with costly-to-maintain original data may lose more value through answer extraction than it gains through discovery. Legal, contractual, or safety constraints can also outweigh referral traffic.
Write the reason down before changing the setting. Then define the evidence that would justify reversing it. A publisher who opts out because traffic is weak needs a different review date from a publisher protecting licensed material. The control is property-level, so one sensitive content type may also reveal a site-architecture problem. Moving that material into a separate, clearly governed property can be more precise than removing an entire publication from generative search.
My verdict: observe first, then exclude with evidence
I would not opt out a healthy publication on day one. I would collect four weeks of page-level and country-level exposure, connect that to conversions as far as the current data allows, then exclude only when the expected loss is acceptable or the content risk is more important than discovery.
This fits a bigger shift we are already tracking. ChatGPT’s European search obligations are making AI discovery more accountable, while ChatGPT’s advertising controls show how quickly answer products are becoming distribution products. Publishers need their own measurement layer before platforms define the value for them.
Read the source and run the test
- Read Google’s worldwide rollout notice.
- Review the current independent rollout coverage.
- Read the preregistered field experiment preprint and keep its study limits attached to any conclusion.
What would make you exclude a page from AI search: weak traffic, loss of control, or a measurable conversion problem?
