Google AI Overviews are easy to over-read. They look like a finished answer, they appear on queries people already care about, and they move often enough that any week can be made to sound like a regime change.
This note is about how to observe them without turning a layout update into a theory of the whole web.
Name the object you are looking at
An AI Overview is a generated module on a search results page. It may include a short synthesis, links or source cards, follow-up questions, and, depending on the query, other elements Google chooses to show. It shares the page with organic results, ads, maps, videos, and other blocks.
That last point is easy to forget if you crop the screenshot. Visibility in an Overview is not the same as visibility in the organic list, and it is not the same as a cited answer in a chat product. A brand can appear in one layer and be absent from another on the same query.
When we record an Overview observation, we record the rest of the visible page as context: whether classic results are still present, whether the Overview is collapsed or prominent, and whether the query looks informational, commercial, local, or navigational.
Query class comes first
Overviews do not behave as one feature. A definitional query, a medical or financial query, a product comparison, and a local service query put different pressures on the system. Some classes invite a short encyclopedia-style answer. Others invite a hedge, a map, or no Overview at all.
If your tracking list is a pile of head terms from a ranking tool, you will mostly see whatever Google is willing to summarize for broad informational search. That can be interesting. It will not tell you how Overviews treat a mid-funnel comparison, and it will not tell you what happens when the query has a city name or a product SKU.
Build the query list the way you would build a prompt panel. Tag each query by job and by commercial intensity. Keep branded navigational queries in their own bucket. “Brand name” queries often produce knowledge-panel behavior and are a weak proxy for category visibility.
Sources on the module are not a complete attribution log
When an Overview shows links, those links are part of the interface. They may correspond closely to the sentences above them, or they may be related pages that Google is willing to offer as next steps. The module does not hand you a footnote trail with the discipline of a journal.
So the coding rules have to be modest:
- Record the URLs or domains that are visible, in the order shown.
- Record whether the brand is named in the generated text.
- Record whether a URL belonging to the brand (or an obvious profile of the brand) appears as a source.
- Do not assume that a source caused a sentence, or that an unsourced sentence came from nowhere.
If you need to talk about “who Google cites,” say “who appeared as a visible source on this query, on this date, in this interface.” That sentence is longer because it is true.
Change is real. Attribution of change is harder.
Overviews expand, shrink, appear on new query classes, and pick up new chrome. Model and retrieval changes can move wording even when the layout looks still. Google also tests. Two people can see different modules on the same query.
A change log is worth keeping. A causal story is worth delaying.
Minimum fields for a change record:
- Query text and query class
- Date, location or language if known, and device
- Whether an Overview appeared
- Length and shape of the generated text (definition, list, steps, comparison)
- Visible sources
- What else occupied the first screen
- A note if the result looked like an experiment (unusual layout, missing blocks)
If wording changes but sources and page composition do not, say that. If the Overview disappears and a featured snippet or organic list remains, say that. “AI Overviews changed” is not a finding until you specify which part moved.
Do not import chat-engine habits wholesale
It is tempting to reuse a ChatGPT or Perplexity codebook without edits. Some event ideas transfer (mention vs citation vs recommendation). The page does not.
Chat products often give you a single column of prose plus optional footnotes. Search gives you a composed page. Users can read the Overview, skip it, or use it as a jumping-off point. Click behavior, when you have it, belongs to a different study. Do not infer engagement from the fact that the module rendered.
Also resist the claim that an Overview citation is “the new page one.” It is one module. Its importance is an empirical question for a query class, not a slogan that can be pasted onto every category.
What a careful Overviews study can still say
Even with those limits, there is real work to do:
- Map which of your query classes currently receive an Overview.
- See whether your entity is named, sourced, mis-attributed, or omitted.
- Compare the Overview’s source list with the organic results on the same page.
- Repeat the map on a fixed query set so you can see coverage move.
Those outputs help a team decide where to improve entity clarity, where to stop expecting a module that is not being shown, and where a chat-engine study is the better instrument.
If you want a query map or a repeated Overviews panel designed as a study rather than a scoreboard, write to [email protected].