Google AI Mode and AI Overviews are not the same surface

These two products share a company, a brand prefix, and a general shape. Almost everything that matters for measurement is different.

The difference in one table

	Google AI Overviews	Google AI Mode
Where it lives	Above the blue links on a normal results page	A separate conversational search experience
When it appears	On some queries only	On every query you ask it
User posture	Typed a short query, deciding whether to scroll	Asked a question, expecting an answer
Answer length	Compressed	Extended, often with follow-ups
Absent answer means	The query did not trigger one, which is real data	Collection failed

That last row is the one that breaks reporting.

An absent AI Overview is data

AI Overviews do not fire on every query. Whether one appears is itself a property of the query, and it is one of the four things worth tracking on this surface. The industry calls it trigger rate.

A tool that treats a missing Overview as a failed collection will either drop that observation or retry it, and both corrupt the denominator. Drop it and your citation rate is computed only over queries that triggered, which silently overstates presence. The honest record is that the query ran, no Overview appeared, and that is what was observed.

refd stores this as an explicit “no answer present” result rather than as an error. It is the difference between “we were not cited” and “there was nothing to be cited in.”

For AI Mode, the opposite holds. It always returns something. An empty result there means collection broke, and treating it as a real absence would understate your visibility.

They are collected differently, and that matters

AI Overviews are part of a search results page, so they are collected the way a results page is collected. AI Mode is a conversational product, collected the way an assistant is.

Two practical consequences:

Different artifacts. AI Mode citation records repeat a full URL inside a domain field, producing a bare site root that duplicates a deeper link from the same origin. Perplexity does the same thing. Counting both inflates citation rate. Google’s citation records also carry favicon and thumbnail service assets in icon fields, which are rendering assets rather than sources.

Different failure modes. A results page can be blocked or return a variant layout. A conversational surface can return a well-formed answer with an entirely different shape than last week. Neither degrades gracefully into the other, which is why merging them into one Google number hides which one broke.

What the market itself says about the gap

Google Search Console does not break out which URLs were cited inside AI Overviews. That absence is the whole reason third-party tracking of this surface exists, and it is worth stating plainly rather than overclaiming past it: no tool, including this one, has access to Google’s own attribution. What a tracker observes is what a collected results page showed at that time, from that collection context.

What to do

1. Track them as two surfaces. Never average them into “Google.”
2. Measure trigger rate on AI Overviews first. A query that never fires an Overview cannot be won, and effort spent on it is effort not spent elsewhere.
3. Expect different winners. The same question can produce different cited domains on each, because the retrieval and the answer format differ.
4. Check that absent Overviews are recorded, not discarded. This is a fast way to find out whether a tool understands the surface.