An AI Overviews experiment is a five-step test: pick a query, record the baseline, apply GEO changes, wait, then record what changed. This article is the template. It's a methodology, not a finished study — I'm not handing you fabricated numbers and calling them proof.
PR teams need a way to test GEO claims instead of taking them on faith. Most advice comes as a list of tactics: add schema, write direct answers, refresh dates. None of it tells you whether the tactic worked on your page, for your query. An experiment does.
Why Run an AI Overviews Experiment
Citation in an AI Overview is binary at any given moment: your content is either quoted or linked, or it is not. That makes it testable in a way rankings rarely are. You do not need a statistically significant sample to see whether a page moved from absent to cited. You need a clean before-state, a defined change, and a clean after-state.
This matters for PR specifically. Coverage and owned content both feed AI Overviews as source material. If a firm can show a client, in concrete before-and-after terms, that a specific structural change produced a specific citation outcome, that is a stronger pitch than a general claim about "AI visibility."
The Five-Step Test Structure
Step 1: Pick a Target Query
Choose a query where the answer is factual, specific, and answerable in two or three sentences. AI Overviews favor queries with a clear, extractable answer over broad or highly debated topics. A query like "what is a press release boilerplate" tests better than a query like "is PR dead."
Pick a query connected to a page you control and can edit. Testing a query you have no content for tells you whether AI Overviews exists for that topic, not whether GEO changes work.
Step 2: Record the Baseline Citation Status
Before touching anything, document the current state:
Does an AI Overview appear for the query at all?
If it appears, is your domain cited, linked, or absent?
If a competitor is cited, what does their answer structure look like: paragraph, list, table, definition box?
What is the exact wording of the AI Overview's summary?
Screenshot this. Save the query text, the date, and the device or location used to check it, since AI Overview results vary by session and region.
Record here: baseline citation status for the target query, before any changes were applied.
Step 3: Apply GEO Changes
Change one page against a documented checklist rather than rewriting the whole site. The changes worth testing in a single pass:
Rewrite the opening paragraph or section to state the direct answer in the first sentence.
Add or correct structured data: FAQPage, HowTo, or Article schema depending on content type.
Break long paragraphs into short, quotable statements a model can extract cleanly.
Update the publish or last-modified date if the content has genuinely been refreshed.
Add a clearly labeled FAQ section with direct question-and-answer pairs.
For the underlying checklist this experiment draws from, see GEO for PR: A Practical Checklist.
Change one variable set at a time if you plan to run this more than once. Changing schema, copy structure, and freshness signals simultaneously tells you the combination worked, not which element mattered.
Step 4: Set a Re-Check Interval
Pick a fixed re-check window before you start, not after you see early movement. A reasonable interval is seven days for the first check and thirty days for a second check, since AI Overview citations can shift as a page gets recrawled and as the model's source pool updates.
Check the same query, in the same way, at each interval. Use the same phrasing, and log location or device settings if they affect results in your testing environment.
Record here: citation status at the 7-day and 30-day re-check intervals.
Step 5: Record What Changed
Document the outcome in the same format as the baseline:
Citation status: absent, linked, or quoted directly.
If quoted, whether the AI Overview now uses language closer to your rewritten answer.
Any change in which competitor sources appear alongside or in place of yours.
Whether the AI Overview box appeared at all, since some queries lose the box over time independent of any page change.
Record here: the specific language or ranking-position change observed in the AI Overview summary after the test period.
What This Test Does and Does Not Prove
A single query test proves correlation between one change and one outcome, on one page, over one time window. It does not prove that the same change will work on a different query, a different domain, or a different topic vertical. Google's AI Overview system also updates independent of any single page, so a citation gain or loss during the test window may reflect factors outside the edit you made.
The way to build confidence is to run the same five-step structure across a set of queries and pages, then look for a pattern rather than a single win. A pattern across ten tests carries more weight than a strong result on one.
Record here: the aggregate citation rate across the full set of tested queries, before versus after.
How to Run This Test Yourself
Set up a simple log before you start: query, date, baseline status, change applied, re-check date, re-check status. A spreadsheet is sufficient. The discipline is in checking the same way every time and resisting the urge to change more than one variable per test cycle.
If your team is also testing visibility in chat-based assistants rather than search AI Overviews specifically, the query structure and answer format matter in similar ways; see How to Rank in ChatGPT Search for how that differs from Google's implementation. And for background on how AI Mode changes the underlying search experience this test is built around, see Google AI Mode Optimization.
For a technical reference on how Google describes AI Overviews sourcing and generation, see Google's own documentation on AI Overviews sourcing. Run the test before you trust anyone's tactics list — yours including.
Frequently Asked Questions
What is an AI Overviews experiment?
It is a structured test that checks whether a page is cited in Google's AI Overview for a target query before a GEO change, applies a defined change, then checks the same query again after a set interval to record what shifted.
How long should you wait before re-checking results?
A first check at seven days and a second check at thirty days is a reasonable structure, since AI Overview citations can change as pages get recrawled and as the underlying source set updates.
Can one test prove a GEO tactic works?
One test on one query and one page shows correlation, not proof. Confidence comes from running the same structure across multiple queries and looking for a consistent pattern rather than relying on a single result.
What should be changed in the GEO test step?
Change one variable set at a time: answer structure, schema markup, or content freshness. Changing several at once makes it impossible to tell which change produced the outcome.
Does losing an AI Overview citation always mean the page got worse?
No. AI Overview boxes can disappear or change for a query independent of any specific page edit, since Google adjusts which queries trigger the feature and which sources it draws from over time.