AI Mode vs. AI Overviews: source delta

Two surfaces, one company

Both surfaces belong to Google, and both reportedly use fan-out. So a lot of coverage treats them as functionally interchangeable, differing only in UI placement: embedded snippet versus dedicated mode.

That may be true. Or AI Mode's more extensive, conversational retrieval could surface a meaningfully different source set than the more compact AI Overview snippet, for the identical query. Nobody has published a direct, same-query comparison yet.

What is known about each, separately

The structural difference behind the question

AI Overviews shows up as a compact answer box, embedded directly above traditional search results. It's generally understood to draw from a fairly narrow, fast retrieval pass, appropriate for a snippet users glance at briefly.

AI Mode is a dedicated, conversational surface. It's documented to use a more extensive query fan-out process, across potentially many more sub-queries per question. If retrieval depth genuinely differs this much between the two, it would be surprising if their cited-source sets turned out identical. The more interesting question is how large that delta actually is, and whether it stays consistent or swings wildly across different kinds of queries.

The delta question

Take an identical query set. Run it through both surfaces, under comparable conditions. What share of cited domains overlap? Where they diverge, what characterizes the difference?

Study design

This uses the same 1,000-query set as the Citation Index, run through both surfaces in the same collection window. Citation sets get compared directly, per query.

A worked hypothetical comparison

Here's what the comparison will look like once real data exists, using invented numbers only. Say the query is "how does a heat pump compare to a gas furnace for cold climates."

AI Overviews cites three sources for it: a manufacturer page, a government energy-efficiency site, and a review publication. AI Mode, for the identical query, cites six sources. One is the same government site. The other five are a forum thread, two comparison articles, and a regional utility company's guidance page.

One source is shared out of eight distinct ones combined. That works out to a Jaccard index of roughly 0.125 for this single query. That's well below the 50% threshold hypothesis MD1 predicts holds on average across the full query set. It also fits MD2's prediction: that AI Mode's retrieval surfaces a meaningfully broader domain set than AI Overviews' more compact snippet.

Pre-registered hypotheses

Why this matters for measurement

Suppose the two surfaces diverge substantially. Tracking "Google AI visibility" as one blended metric, which many current tools do, would then hide meaningfully different performance on the two products. That's the same measurement failure the measurement standard argues against more generally.

Why the delta likely isn't uniform across query types

A single blended overlap figure, across the whole query set, would risk hiding real structure. A narrow, factual query, like a specific unit conversion or a well-established definition, plausibly produces similar citations on both surfaces regardless of retrieval depth. There's simply less room for divergence when few sources are authoritative to begin with.

A broad, comparative, or current-events-sensitive query is different. There, AI Mode's deeper fan-out has more room to surface extra sub-queries and extra sources. That kind of query would plausibly show a much larger delta. Reporting overlap broken down by query type, rather than only as one overall average, is planned specifically to avoid hiding this kind of variation.

What to do about it now

You don't need to wait for this study to hedge sensibly. Write for AI Overviews' narrower snippet first. State your clearest, single best answer near the top of the page. Use plain language a fast retrieval pass can lift easily.

Separately, support AI Mode's deeper, conversational retrieval too. Cover realistic follow-up questions later on the same page. Its fan-out process is built to chase exactly that kind of sub-query. One well-structured page can reasonably serve both surfaces at once. It just needs a strong top-line answer, plus genuine depth underneath it.

Limitations

• Both surfaces are actively developed and could converge or diverge further after this study runs — a snapshot, not a permanent architectural fact.
• Overlap measures agreement, not which surface's citations are better — that would require a separate quality-evaluation method this study doesn't attempt.