A page can appear in an AI answer without generating a visit. A visit can happen without producing an order. And an increase in orders can have causes unrelated to search. Keeping those situations separate is the first step toward using visibility data without turning it into a revenue promise.
Google announced dedicated reports for generative features in Search and Discover, with worldwide availability confirmed on August 31, 2026. The announcement lists impressions, pages, countries, dates and, for Search, devices. The data remains part of the overall reports: it is not additional volume to add on top. Google’s official announcement.
For an ecommerce team, the useful question is not simply “do we show up in AI?”. It is which pages gain exposure, in which markets, how consistently, and what deserves investigation. This guide proposes a working method with a fictional example and verifiable decisions. It does not report results from an actual AGTI or client account.
1. Define the question before opening the chart
Start with a decision the report can help inform. “We need a stronger AI presence” is too broad. “Which installation guides started appearing in the market we serve?” gives you a reason to select particular pages, dates and countries.
Distinguish three content groups: product pages, category pages and educational resources. A guide may answer a question early in the buying process; a product page may help someone already comparing specifications. Judging both only by immediate orders creates an unfair comparison.
Create a short list of priority URLs and assign a purpose to each. Do not select only high-volume pages: include commercially important pages and recently improved resources. This list becomes a reference for reviewing the data, not a guarantee that every URL will appear in the report.
2. Read the dimensions without inventing metrics
Use documented dimensions to narrow an investigation, not to automatically explain customer behavior. A growing page may have benefited from seasonal demand, a content change, a different set of answers or several factors together.
Before calculating an indicator, check which fields your interface actually offers and what each one means. Do not create an “AI conversion rate” by dividing store orders by impressions. Those orders may have originated from advertising, direct visits, campaigns or other organic traffic.
Likewise, a displayed URL does not, on its own, reveal the full question someone asked, their buying intent or why Google selected that page. Record those gaps as analytical limits. An explicitly incomplete hypothesis is better than a precise-looking conclusion built from incompatible data.
3. Build a comparison you can repeat
Use the same property, surface, page set and filters for both windows. Search and Discover should not be treated as interchangeable discovery contexts.
Prefer periods of equal length with a similar mix of weekdays. If the available history is still short, establish a baseline and allow observations to accumulate. Do not fill missing weeks with zeroes or interpret unavailable history as a performance decline.
Keep a working sheet containing the review date, reporting interval, filters, included pages, observed value and relevant site changes. You can maintain it manually; the method does not depend on a particular export option being available.
Maintain an intervention log as well. New articles, downtime, URL changes, catalog updates and campaigns belong alongside the numbers. The log does not prove causation, but it helps prevent interpretations detached from what actually happened in the business.
4. A fictional example: 40% growth can conceal a relevance problem
Consider a Brazilian store selling equipment for small workshops. For the same page set and identical filters, two 28-day windows show:
| Segment | Period A | Period B | Absolute change |
|---|---|---|---|
| Impressions in the complete selection | 10,000 | 14,000 | +4,000 |
| Brazil | 7,000 | 7,700 | +700 |
| Other countries | 3,000 | 6,300 | +3,300 |
These are teaching figures, deliberately constructed so the country groups do not overlap. Total growth is 40%, but growth in Brazil is 10%. Brazil’s share moves from 70% to 55%. If the store serves Brazil alone, most additional exposure occurred outside its current serviceable market.
That does not make international visibility bad. It could reveal interest in a tutorial or a future opportunity. It simply prevents the team from describing the 40% increase as an equivalent expansion in its present commercial opportunity. The same reasoning applies to any business serving a defined country or region.
Now suppose a separate website analytics measurement records 500 and 520 organic sessions in the two windows. That is 20 additional sessions, or 4%. Because this measurement does not necessarily isolate visits from generative features, it cannot be used with the 14,000 impressions to calculate a click-through rate. The numerator and denominator do not represent the same population.
A responsible conclusion would be: “Exposure increased, mainly outside our current market. We will identify the content involved and investigate whether this represents a commercial opportunity or informational reach.” It would not be: “AI increased our sales by 40%.”
5. Investigate pages in stable groups
After the overview, look for changes within coherent sets. For example, separate maintenance guides, machinery categories and accessory product pages. Compare each group with itself before comparing it with another group.
If a guide rises while a category declines, inspect purpose and content before deciding that one format has replaced the other. The guide may address a technical question the category was never intended to answer. A decline concentrated in changed URLs also calls for a different investigation from a sitewide decline.
Avoid applying so many dimensions at once that only a handful of observations remain. Moving from two impressions to six is a 200% increase, but it is still a small volume. Always show absolute numbers alongside percentages, and avoid a universal threshold for “good performance”.
For multilingual pages, check editorial equivalence. Does the English guide cover the same subject at comparable depth? Does it serve a market where the business can sell or deliver services? Visibility across languages should not be consolidated without explaining the purpose of that combined view.
6. What to review before buying “AI optimization”
Google still points to established SEO foundations: a page must be indexed and eligible for a snippet, with no inclusion guarantee. No special AI file or exclusive markup is required. Official guidance for AI features.
Turn that principle into a concrete content review. On an equipment page, can a visitor identify power requirements, dimensions, intended applications, limitations and included parts? Repeating “quality and performance” does not answer those questions.
For a compatibility guide, document the criteria behind the advice. Explain which dimensions or versions must match, which situations require confirmation, and when the recommendation no longer applies. If evidence does not support a statement, remove the promise or state the uncertainty clearly.
These are editorial and operational improvement proposals, not secret selection factors. The objective is a useful, verifiable page for anyone who reaches it, regardless of the search format that brought them there.
7. Prioritize an intervention you can evaluate
Choose a small group of pages with an identifiable problem. Instead of rewriting the whole catalog, start with product pages where a missing measurement repeatedly generates support questions. Record the original situation, the information added and the publication date.
Decide in advance what you will monitor. Alongside exposure, the team might observe questions about that attribute, navigation to compatible products or requests for a quote. Keep each observation tied to its source; do not blend support, analytics and Search Console into a single metric.
When reviewing the outcome, check for other changes in the period. Pricing, stock, campaigns and site availability may interfere. A before-and-after comparison is useful for monitoring, but it is not a controlled experiment.
If the content becomes more accurate and reduces uncertainty, that can already justify the improvement. You do not need to credit an algorithm for every benefit or demand an immediate increase in impressions after every update.
8. Separate visibility, behavior and business outcomes
Organize monitoring into three layers:
| Layer | Management question | Main caution |
|---|---|---|
| Visibility | Where does our content appear? | Exposure is not a visit. |
| Behavior | What do visitors do on the site? | Not every organic session originates in a generative feature. |
| Business | Are there useful inquiries or orders? | Do not claim causation without evidence. |
A page may serve support, discovery or comparison. Define that role before expecting a particular commercial outcome. Consider operational capacity as well: attracting inquiries from unserved regions can increase work without creating viable revenue.
In meetings, present one conclusion and one next action for each problem. “The guide gains exposure but leaves compatibility unclear; we will add a table validated by the technical team” is more useful than “we need better GEO”.
9. A lightweight routine that does not chase fluctuations
A short weekly review can record the overview; a less frequent analysis can evaluate interventions. Match the cadence to data volume and stability rather than treating daily movement as an urgent content request.
During each review, ask: has volume changed? Has the mix of markets or pages changed? Is there a technical or editorial change to investigate? Which action has an owner and a clear completion criterion?
If evidence is insufficient, record “monitor”. Not taking immediate action can be the correct decision. Rushing out content in response to a single point on a chart tends to create rework and weaken catalog quality.
Conclusion
The new report provides another way to observe presence in Google’s generative features. Its practical value depends on well-defined questions, comparable segments and explicit limits.
Start with a few pages, preserve the filters and document changes. Use the data to prioritize content and technical investigation without assuming that impressions translate into sales.
AGTI can help connect content, technical structure and ecommerce operations in a plan of verifiable improvements. Contact the team to discuss the problem your store needs to solve.