I tell clients to open every major section with a self-contained answer of about 40 to 60 words, because it reads better for a skimming human and because an AI system quoting the page gets a clean passage rather than half a thought. I have never tested whether the second half of that claim is true. This is me doing that.
The question
Does rewriting the opening of a page into a short, self-contained answer change how often that page gets cited in Google’s AI Overviews, or mentioned by AI assistants, compared with a matched page left as it was? The claim under test is my own advice, which is the reason it is worth testing.
Why this matters
I say this to clients, it appears on my own service pages, and it shapes how content gets written across every site I work on. If it does nothing measurable for citation, that advice needs downgrading to a readability recommendation rather than an AI search one. Testing your own recommendations is uncomfortable and it is the point.
The method
Twenty pages across sites I control, split into two matched groups of ten by topic, length and current visibility. Group A gets its opening rewritten into a 40 to 60 word self-contained answer that names the subject and carries the single most useful fact. Group B is left alone.
| Element | Setting |
|---|---|
| Sample | 20 pages, 10 treated and 10 matched controls |
| Change | Opening passage only. No other edit, no new sections, no new links |
| Baseline | Citation capture for every page before any change |
| Measurement | Same query set run weekly on Google AI Overviews, ChatGPT and Perplexity |
| Repeats | Each query run three times per capture, because AI answers vary |
| Duration | 10 weeks after the change is published |
Only the opening changes. The temptation is to improve the whole page while you are in there, and that would make the result meaningless.
What I expect to happen
My prediction, written before the result: a small increase in citation for the treated group, clear enough to notice and too small to be conclusive at this sample size. I expect the effect to be larger on question-shaped queries than on commercial ones.
I also expect a lot of noise. AI answers vary between runs, which is exactly why each query gets run three times per capture and why the baseline matters.
Results
The result is pending and stays that way until the ten weeks are up. When it fills, it will carry the baseline and final citation counts for both groups, the per-platform split, and the variance across the three runs of each query, so you can see how noisy the measurement was rather than just the headline.
[PAUL TO ADD: the figures, once the test has actually run.]
What I concluded
The conclusion is pending and gets written from the data. If the treated group shows no difference, that goes up plainly, and my own service pages get amended to match. Publishing a test that undermines your own advice is the only version of this exercise worth doing.
How to repeat this
Running this yourself needs a set of pages you control, a fixed query list and the discipline to change one thing. It takes about a day to set up and ten weeks to answer, and the step people skip is the baseline, which makes the whole thing unreadable afterwards.
- Pick twenty pages of similar type and visibility. Split them into two matched groups.
- Capture a baseline: run your query set against each platform three times and record who gets cited.
- Rewrite only the opening of group A into a 40 to 60 word self-contained answer. Publish.
- Leave group B completely alone, including no unrelated tidying.
- Recapture weekly on the same queries, at the same time of day, for ten weeks.
Limitations
Twenty pages on sites I control is a small, unrepresentative sample, and AI citation is influenced by things this test does not hold constant, including site authority and competitor changes. A positive result would be a signal worth acting on. It would not be proof that the passage caused it.
FAQ
What is an answer passage?
An answer passage is a short, self-contained paragraph at the start of a section that answers the question the heading implies. It names its subject rather than saying “it”, includes the most useful fact, and makes sense if quoted on its own away from the page.
Has Google said answer passages help with AI Overviews?
No, and that is worth being clear about. Google says there are no additional requirements or special optimisations for AI Overviews beyond ordinary SEO. The 40 to 60 word guide is a working practice of mine, which is precisely why it deserves testing rather than repeating.
Why run each query three times?
Because AI answers vary. The same prompt can return different sources minutes apart, and on a single capture you cannot tell a real change from normal variation. Three runs per capture gives at least a rough sense of how noisy the measurement is.
Will you change your advice if this fails?
Yes, and the service pages that carry the advice get amended too. Advice that survives only because nobody checked it is not worth giving. If the effect turns out to be readability rather than citation, that is still useful and it gets described that way.
Sources
- Google Search Central, AI features documentation, stating there are no additional requirements to appear in AI Overviews or AI Mode beyond standard SEO. Accessed 16 September 2026.
- Google Search Central, guide to optimising for generative AI features, on non-commodity content and foundational SEO. Accessed 16 September 2026.
Want your pages checked for this?
Whether your content is quotable is one of the few parts of AI search you can actually control, alongside crawler access and how clearly your business is described. It takes a read of your key pages rather than a tool to judge properly.
Get in touch, or start with a free assessment. More tests are on the tests hub.