How to track Google AI Overviews and AI Mode: in our panel, a third of AI Mode answers cited nothing

What we measured

Promvia, which publishes this guide, is our product, and the numbers come from its public AI Source Index: 40 fixed English buyer questions (16 "best X", 8 "alternatives to X", 8 comparisons, 8 "how do I choose"), asked once per run on Google AI Overviews and Google AI Mode, among other engines, on 3, 7, 14, 21 and 28 September and 5 October 2026. That is 240 checks per surface; one AI Overviews check failed on 5 October and is left out, so AI Overviews has 239.

• A domain is the lowercase host without a leading www., counted once per answer. Google's own search and shopping links are not counted as a source, as in the public index.
• "Cited nothing" here is an answer that named no domain outside Google: either no link at all, or only Google's own links.
• Every rate is a count out of a count, with a 95% interval in the published summary. Every number in this guide was computed twice, by a Python script and an independent Node script, and the two agree on all 34 values.

What to track for each question

A rank tracker records one number per keyword. An AI answer needs a few more fields, because most of what goes wrong in AI tracking is reading one state as another:

• Whether Google showed an AI answer at all. No AI Overview on the page is a different fact from an AI Overview that left you out.
• Whether the answer cited any source. An answer with text and no links has no citation slot at all.
• Which domains it cited. This is your competitive set for that question, and in our panel three of every four domains Google's AI cited were not on its organic page one.
• Whether your own domain is among them, and whether the answer text names your brand. The two can differ: an answer can name you and link a review site.
• Your organic position for the same query, from the same results page. It is the comparison that tells you whether the AI answer follows the ranking or not.
• Whether the check itself failed. A provider error is missing data, never "not cited".

Track AI Overviews and AI Mode as two separate engines. They share a brand, not a source list: in our panel only 17.4% of the domains they cited for the same question in the same week were cited by both (see below).

Branded and unbranded questions

Most of your list should be unbranded questions, the ones a buyer asks before they know your name: "best X for Y", "alternatives to Z", "how do I choose X". Those are where the answer decides which sites are considered at all, and they are what our panel asks.

Keep a small set of branded questions too, such as "is your brand any good" or "your brand vs a competitor", to see what Google's AI says about you and which pages it quotes when it does. Read them differently: your name is in the question, so the answer naming you says little. The useful signal in a branded answer is which domains it cites and what it says about you.

How often to check, and why one run is noise

Google's AI answers change from week to week even when nothing on your side changed. In our panel, the same surface cited the same domain for the same question one week later less than half the time:

• AI Overviews kept 502 of 1,123 cited domains from one weekly run to the next (44.7%).
• AI Mode kept 185 of 462 (40.0%). Counting only week pairs where both answers cited something, 185 of 377 (49.1%): part of AI Mode's turnover is answers that cited nothing the next week.
• Whether an answer cites anything at all also moves. Of the 40 questions, 25 got an AI Mode answer with no website in at least one of the six runs, but only 4 in all six; 21 switched between citing sources and citing none. For AI Overviews, 18 of 39 questions switched, and none was empty every week.

So check at an interval that gives you several readings before you decide anything, and decide on the pattern, not the latest check. A weekly check of a fixed list is the minimum that gives you a trend; more frequent checks give you more readings per decision, not a truer single reading. Before treating a drop as a loss, wait for the next run: in the full study a quarter of the domains that dropped out of an answer were back the following week (how long AI citations last).

AI Overviews and AI Mode cite differently

AI Overviews cites more sites per answer. Among answers that cited at least one domain, the median AI Overview cited 6 domains (1,323 across 204 answers, a mean of 6.5) and the median AI Mode answer cited 3 (596 across 163, a mean of 3.7).

They also cite different sites. In 239 question-and-run pairs where both surfaces returned a check, both cited something 140 times, only AI Overviews 64 times, only AI Mode 22 times and neither 13 times. Within the 140 pairs where both cited something, 206 of the 498 domains AI Mode cited were also cited by AI Overviews for the same question that week (41.4%), and 206 of the 890 AI Overviews domains were also in AI Mode (23.1%). Taken together, 206 of 1,182 distinct domains were cited by both: 17.4%. Most pairs shared at least one domain (116 of 140, 82.9%).

Ahrefs counts URLs and we count domains, so the numbers are not the same measure, but they point the same way: being cited in one of Google's AI surfaces tells you little about the other. Look at each one's own source list on the AI Overviews and AI Mode pages of the index.

"No AI answer" and "answered with no sources"

An empty AI Mode result is common in our panel, and it is not one thing. Of AI Mode's 77 answers without a website, 51 had no link at all (21.3% of 240) and 26 linked only to Google's own search or shopping pages (10.8%). AI Overviews never had the second kind: all 35 of its empty results had no link at all.

The public rows do not say whether a result with no link was "Google showed no AI answer" or "an answer with no sources". For two runs we can tell: on 28 September and 5 October, for 38 of the questions, AI Mode gave an answer with no sources 22 times in 76 checks and showed no answer 4 times, and AI Overviews gave 8 source-less answers in 75 (Google's AI cites pages Google doesn't rank). Most of AI Mode's empty results are answers written without links, not missing answers.

Our 32.1% is far above both. The studies differ in when they ran (mid and late 2025, against September and October 2026 for ours), in what they asked (keyword sets against our 40 long buyer questions) and in how they collected the answer (our checks go through a SERP data provider). We cannot tell from our data which of these explains the gap, and we cannot rule out that some of our empty AI Mode results are the provider returning an answer without its links. Read our figure as what our collector saw on these questions, not as a share for AI Mode in general.

For your own tracking, the rule is the same either way: a question where Google cited nothing is not one you lost. Nobody won that answer. Count it separately, keep it out of your citation rate's denominator or report it beside it, and see whether it stays empty over several runs before you decide the question has no citation to compete for.

Read citations as a list, not a rank

A citation in an AI answer is not a position on a results page, and the two lists overlap less than most people expect. In our study of the same 40 questions, 140 of 594 domains cited by AI Overviews and AI Mode were also on Google's organic page one for the same question (23.6%), and only 13.5% of page-one domains were cited (140 of 1,038). So report citations as which sites the answer used, and whether you were one of them, rather than as a number you move up and down. Keep your organic rank beside it as a separate column, not a stand-in.

How Promvia collects AI Overviews and AI Mode

• Both are read from Google's own results data through DataForSEO's SERP API, not through a chat API. AI Overviews comes from Google's organic results page, with the option that also loads an AI Overview Google defers; AI Mode comes from DataForSEO's dedicated AI Mode endpoint.
• Each check is one request, read once: we do not sample these two surfaces twice the way we do the chat engines. Location and language follow the market you track; the index's global panel uses Google US in English.
• The AI Overviews request returns the organic results too, so your organic position for the same question is recorded from the same page at no extra cost.
• "No AI answer", "not cited" and "check failed" are three different states in the product, and a gap in data is never shown as a loss. The full collector notes are on the methodology page.
• AI Overviews and AI Mode are among the default engines on Pro and Agency, which by default check every tracked question on them automatically. Starter's automatic checks run on ChatGPT, Perplexity and Gemini by default, and every paid plan can check all the default engines on demand. The current schedule per plan is on the pricing page.

Tracking a surface tells you where you stand; it does not decide whether Google cites you. For which kinds of sites the engines cite, see what AI engines actually cite.

Limits

• One fixed panel: 40 English buyer questions, Google US. Not a sample of all questions, languages or markets.
• One sample per surface per question per run, so a single change can be the surface's own variation.
• Six runs, one week apart (the first four days apart). Nothing here sees what happens within a week.
• Results read through a SERP data provider, not a logged-in browser. Personalised results may differ.
• The no-answer and no-sources split is known for two runs only.

Frequently asked questions