Google AI Mode and AI Overviews Share at Least 363 of the Top 500 Sources — and Why We Withdrew the Gap Number

Search and brand teams checking AI visibility on a single Google surface are checking half the picture, even within Google. Machine Relations Index data covering the 500 most-cited domains across six answer engines shows Google AI Mode and Google AI Overviews — both built on the same Gemini foundation — citing a shared floor of 363 of those domains and diverging on at most 95. How much of that divergence is Google and how much is our own instrument is the subject of the correction immediately below, which is the most useful thing on this page.

Correction, September 26, 2026 — the headline number is withdrawn

This page was published on September 24, 2026 with the headline "67 Domains Apart," reporting that Google AI Mode reached 444 of the top 500 cited domains and Google AI Overviews reached 377. Both figures are floors, not measurements, and the 67-domain gap between them is withdrawn.

The Machine Relations Index — our own instrument — had dated collection failures on both Google surfaces inside this page's measurement window. It recorded zero citations from Google AI Overviews on any monitored query between August 15 and September 24, 2026, 41 of the window's 132 days; and zero from Google AI Mode between September 14 and September 24, 11 days. A third arm, Gemini, has 22 intermittent blind days in the same window. The cause on the two Google lanes was a defect in the Index's collection path, since fixed: the prior collector graded a response as a success whether or not it carried content.

A presence count only accumulates. A blind day can subtract a domain from an engine's observed set and can never add one. So every figure on this page for AI Mode, AI Overviews and Gemini is a lower bound, and the arithmetic that follows is unavoidable: the surface with more blind days — AI Overviews, at 41 — is the one whose count is most depressed, which is precisely the direction that manufactured the gap this page was built on. A 67-domain difference computed between two differently-depressed floors is not a finding about Google. It is a finding about our collection.

What survives, and why the page is still here. Presence is not symmetric with absence. A domain recorded as cited by both surfaces was genuinely cited by both — collection gaps cannot invent a citation. So the 363 domains this page found in both Google surfaces' citations are a hard floor: the two surfaces share at least 363 of the top 500 cited domains, and may share more. The 95 domains recorded as divergent are a ceiling: at most 95, and an unknown share of that is our own blind days. The operational advice at the foot of this page is unchanged and, if anything, better demonstrated now than when it was written — one surface's silence is not proof of absence, and this page is the worked example.

Key takeaways

• Google AI Mode and Google AI Overviews are recorded citing the same 363 of the top 500 domains. That is a floor: both surfaces genuinely cite all 363, and the true shared set is at least that large.
• The 95 domains recorded as divergent between the two surfaces are a ceiling, not a count. Our instrument was blind on AI Overviews for 41 of 132 days and on AI Mode for 11, and both gaps push in the direction of manufactured divergence.
• Of the six engines, ChatGPT, Claude and Perplexity collected without a gap in this window; one Claude date was flagged by the same scan and carries exactly one answer run, a scan artifact rather than a gap. ChatGPT's 361 of 500 (72.2%) is exact and is the narrowest head overlap of the six — every other arm's floor sits above it, so that ranking holds in the one direction a floor can carry.
• TechRadar.com is genuinely a head domain: rank 7 of 23,280, confidence grade A, cited in 460 of 16,475 observed runs by ChatGPT, Claude, Gemini and Perplexity. It was not observed in either Google surface's citations in this window, but with both lanes carrying blind days that observation cannot be separated from collection and is not published here as an absence.
• The finding is a per-engine citation pattern, not a ranking signal or a causal claim about why any one domain is or is not cited by either surface.

What Machine Relations measured

The Machine Relations Index tracks 23,280 cited source domains across 15,782 to 16,475 answer runs (the run count varies by cut) over a 130-day window ending 2026-09-23, split across six engines: ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity. Each of the top 500 ranked domains carries an engine_breadth field recording which of the six engines cited it at least once in the window — not a rate, a presence flag. That field is what this piece reads engine by engine (Machine Relations Index, release generated 2026-09-23, window 2026-05-10 to 2026-09-23).

Reading engine breadth across the top 500 domains gives a head-overlap table. It is not a ranking of the six, and the September 26 correction above is why: three of the six rows are floors.

Engine	Domains reached (of top 500)	Share	Exact or floor
Perplexity	490	98.0%	Exact — collected without a gap
Claude	463	92.6%	Exact — collected without a gap
Gemini	452	90.4%	Floor — 22 blind days
Google AI Mode	444	88.8%	Floor — 11 blind days
Google AI Overviews	377	75.4%	Floor — 41 blind days
ChatGPT	361	72.2%	Exact — collected without a gap

Two readings survive this table and no more. Perplexity and Claude sit within a few points of universal head coverage, and both counts are exact. And ChatGPT, the most consumer-recognized name on the list, has the narrowest reach of the six: its 361 is exact, and every other arm's floor sits above it, so no amount of recovered collection moves it off the bottom. That is a fact that matters for any brand team whose AI-visibility check is a single ChatGPT prompt.

What this table cannot support is an order among the three floored rows, or a claim that Perplexity's 490 is the widest of the six — a floor of 452 or 444 could sit anywhere above itself, including above 490. The rows are reported rather than ranked.

Two Google surfaces, two different citation pools

The gap that carries the most operational weight is inside Google. AI Mode and AI Overviews are commonly treated as one product with two skins: both run on Google's query fan-out technique and a custom Gemini build, both surface inside Google Search, and Google itself reports both operating at billions-of-users scale — AI Overviews alone reaches roughly 2.5 billion monthly users (nadiamohamed.me, citing Google I/O 2026). But the products differ in how a user reaches them — AI Overviews appear automatically above organic results; AI Mode is a distinct tab a user opens for multi-turn, conversational search — and the Index data shows the citation pools differ too, not just the interface.

• 363 domains are cited by both. This is the one solid number in the set: a recorded citation cannot be an artifact of a blind day, so both surfaces genuinely cite all 363, and the true shared set is at least that large.
• At most 81 of the top 500 appear in AI Mode's citations but not AI Overviews' — PCMag, Capterra, GlobeNewswire, Fortune and Fool.com among those recorded. AI Overviews was blind for 41 of the window's days, so an unknown share of these 81 is our collection rather than Google's behaviour.
• At most 14 of the top 500 appear in AI Overviews' citations but not AI Mode's: Mayo Clinic, Kiplinger and OWASP among them. The same caveat applies in the other direction, at 11 blind days.
• At most 42 domains in the top 500 are cited by neither Google surface. Each one is reachable by other engines; whether it is genuinely invisible to both Google products, or merely unobserved by us on the days they cited it, this release cannot say.

A brand tracked only through AI Overviews and a brand tracked only through AI Mode may be reading two different citation graphs, even though both graphs are labeled "Google." The shared floor of 363 says the overlap is large; the divergence ceiling of 95 says how much room is left for a real difference, and our own blind days sit inside that room.

The TechRadar case

TechRadar.com is the seventh most-cited domain in the entire 23,280-domain index, at confidence grade A: cited in 460 of 16,475 observed runs, on 101 of 130 observed days. Those figures are exact. It is cited by ChatGPT, Claude, Gemini and Perplexity — and the first two of those are exactly measured arms, so its head status is not in question.

Neither Google AI Mode nor Google AI Overviews was observed citing it during the window. This page originally published that as a fact about Google. It is not one, and the withdrawal is instructive: with AI Overviews blind for 41 of 132 days and AI Mode for 11, an absence in our record is exactly what a collection gap looks like from the outside. TechRadar may or may not be absent from Google's AI citations; this release cannot tell the two apart, and neither can any other instrument that does not publish its own blind days.

That is the real lesson of this case, and it is a stronger one than the page started with. A search team validating AI citation performance by checking only Google's AI surfaces, on a handful of days, would conclude TechRadar has no AI presence at all — the same error our own instrument made, with 130 days of data and a pipeline built for the job. Absence of observation is not observation of absence, whoever is doing the observing.

The same caution runs the other way. Mayo Clinic and Kiplinger were recorded in AI Overviews' top-500 citations but not AI Mode's — a difference a team checking only AI Mode would miss, if it is a difference at all.

Why the two surfaces might diverge

The Index data records presence, not cause, so this is a bounded, sourced note rather than a claim about Google's ranking mechanics. Public documentation on Google's crawler controls describes Google-Extended as the token governing whether a site opts into Gemini's model-training and grounding uses, distinct from the Googlebot token that governs standard Search inclusion — meaning a site's robots.txt posture can, in principle, affect one Google AI surface's access to a page without affecting the other or classic Search (capconvert.com). Whether crawler-access differences, live-retrieval differences between a passive summary and an active multi-turn session, or query-mix differences between the two surfaces explain any part of the recorded divergence is a question the engine-breadth field alone cannot answer. Nor, after the correction above, can this release establish how much divergence there is to explain: at most 95 of the top 500, minus whatever share our own blind days produced. The mechanism note stays because it remains the most plausible published account of how two Google surfaces could differ at all; it is no longer offered as an explanation of a specific measured gap.

What this changes for a search team

A single-surface AI visibility check — one ChatGPT prompt, one AI Overviews screenshot — is a sample of one engine's citation pool, not a market read. The practical implications:

• Checking AI Overviews tells a team nothing reliable about AI Mode citation status, and vice versa; treat them as two engines, not one line item. The two surfaces share at least 363 of the top 500, so the overlap is real — but the remainder is unresolved, not zero.
• A "not cited by Google AI" conclusion needs both Google surfaces checked, on enough days to survive a bad one, before it is safe to publish or act on internally.
• ChatGPT's narrower head overlap (72.2% of the top 500, exact) means a domain's ChatGPT-citation status is the least representative single-engine signal of the six for judging overall AI citation reach.
• Perplexity and Claude's near-universal head overlap (98.0% and 92.6%, both exact) makes either one, alone, a reasonable proxy for "is this domain in the general citation pool" — but neither substitutes for a same-surface Google read.
• Ask any AI-visibility vendor, including this one, which days its collection failed. An instrument that cannot answer cannot tell you whether an absence in its dashboard is the engine's behaviour or its own. Ours is published above.

Method note

Figures are read directly from the Machine Relations Index's public release (machinerelations.ai/data/machine-relations-index.json, generated 2026-09-23, window 2026-05-10 to 2026-09-23, 130 days observed, 23,280 cited domains). The head-overlap table covers the 500 domains carrying a populated rank field — the only domains for which the release publishes a head ranking; the remaining 22,780 domains are ranked null in the public view and are outside this analysis. engine_breadth.engines records whether an engine cited a domain at least once in the window; it is not a citation rate and not a ranking weight. TechRadar's per-run and per-day figures are read from its own mri_score_v2.overall record in the same release. No Answer-Source Fidelity rate is used in this piece: Fidelity currently reports per-engine supported rates for ChatGPT, Claude, Gemini and Perplexity only, and does not yet cover either Google surface, so it cannot answer this question.

FAQ

Are Google AI Mode and Google AI Overviews the same product? No. They share underlying technology — Google's query fan-out approach and a Gemini model — but are different surfaces: AI Overviews are a passive summary layered on standard search results, while AI Mode is a separate, actively opened, multi-turn conversational experience.

Does a domain being absent from Google AI Overviews mean it has no AI citations at all? No — and it may not even mean it is absent from AI Overviews. TechRadar.com is cited by ChatGPT, Claude, Gemini and Perplexity and ranks 7th of 23,280 cited domains overall, while not being observed in either Google surface's citations in this window. Because our collection was blind on AI Overviews for 41 of the window's 132 days and on AI Mode for 11, that non-observation is not published as an absence.

Which engine has the widest reach into the most-cited domains? This release cannot say. Perplexity reaches 490 of the top 500 (98.0%) and that figure is exact, but three arms — Gemini, AI Mode, AI Overviews — carry floors that could sit anywhere above their recorded counts. What the release does settle is the bottom: ChatGPT's 361 of 500 (72.2%) is exact and every other arm's floor sits above it.

Did this page change after publication? Yes. It was published on September 24, 2026 claiming a 67-domain reach gap between the two Google surfaces. On September 26 that claim was withdrawn, in full and at the top of the page, after we found dated collection gaps in our own instrument on both surfaces inside the measurement window. The figures were never re-run to look better; the shared-source floor of 363 is from the same release as the withdrawn gap.

Does this data show why the two Google surfaces cite different domains? Not directly. The Index records which domains each engine cited, not why. Public documentation on Google's crawler-access controls (Google-Extended) describes one plausible mechanism, but confirming it would require access data Machine Relations does not publish.

Where does this data come from? The Machine Relations Index, an editorially independent research instrument from the team behind Paralax, refreshed on a rolling basis and covering six answer engines: ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity.