What AI Overviews Really Cite: Data from Our Multi-Engine Tracker
We monitored AI responses on a set of real queries between July and August 2026 using a multi-engine tracker: which sources Google AI Overview, ChatGPT and other engines cite, how much weight video and social carry, what happens to small brands, and why the same question gets different answers from different engines. Verifiable observations, with declared dates and sample limitations.
What AI Overviews really cite: data from our tracker
Verifiable observations from multi-engine monitoring: Google AI Overview, ChatGPT and other engines compared on the same queries.
Between July and August 2026, we tracked AI responses to a set of real queries using the integrated monitoring system in AraLith COS, saving for each check the generated response, the cited sources and the organic positions. This article collects the observed patterns: some confirm what the GEO literature has been saying for months, others do not.
In summary: what AI Overviews cite
AI Overviews primarily cite YouTube videos and social content for informational queries, English-language sources even for queries written in Italian, and present risks of identity confusion for minor brands. For some price-related questions, Google builds the AI response directly from product pages rather than editorial articles. These observations emerge from the data collected by the AraLith multi-engine tracker on a sample of real queries between July and August 2026.
What does the AI Overview cite?
The sources cited by AI Overviews do not always match the top organic results. In our AI Overview tracker data (7 queries, snapshots from July 21 to August 17, 2026) the clearest patterns are four:
- For some informational queries, Google anchors the answer to videos and social media (YouTube Shorts, TikTok, LinkedIn), not articles.
- AI Overview citations and those from ChatGPT or Perplexity diverge: same question, different sources, sometimes in different languages.
- Small brands risk identity confusion with namesakes if the entity is not well established across external sources.
- AI Overview data in Italy should be read with caution: local SERPs and neutral SERPs tell different stories, and answers change over time.
How was this data collected? Method and sample limitations
How monitoring works: for each tracked query, the system periodically queries multiple engines (Google with AI Overview and conversational AI engines) and saves a snapshot with the full response, the list of cited sources with snippets, the status of target content, and organic positions in the same SERP.
The sample for this article: 7 tracked queries related to our industry (software, SEO, AI content), with snapshots collected between July 21 and August 17, 2026. This is a small, vertical sample: the observations that follow are patterns documented on our own queries, not statistics generalizable to the entire web — and we state this explicitly because it is the same honesty we ask of the sources we cite.
Each observation below includes the date of detection. AI responses are volatile: the same query can receive different answers and sources within hours.
Observation 1: for some queries Google still refers to videos and social media, not articles
The clearest case in the sample: for the query "AI overview optimization" (Google detection on August 17, 2026), the AI Overview cited 7 sources. The sources anchored to the response passages — those from which the text is actually built — were 3 YouTube Shorts, 1 TikTok video and 1 LinkedIn post. The two traditional web guides present appeared only as source cards, with no anchoring to any text block.
The operational reading: for this informational query the model built consensus from video and social content, not from articles — including those from specialist sites. Those who cover only the article format compete for the side cards, not for the response text. This is a pattern observed on a specific query, not a general rule: but it is the type of information that no generic best practice gives you, and that changes editorial priorities (a short video derived from a guide becomes a GEO asset, not an accessory piece of content).
Observation 2: the same Italian question receives different answers (and in different languages) from different engines
On the same query "ai overview optimization", within the same short time window (August 17, 2026), multi-engine checks produced three radically different answers:
- Google responded in Italian, with Italian sources (videos, social media and two industry guides).
- A conversational AI engine responded on the right topic but citing almost exclusively English-language SEO industry sources (Search Engine Land, SE Ranking, Semrush).
- Another engine interpreted the query as machine learning optimization, citing arxiv.org, TensorFlow and Optuna: "ai" read as English changed the meaning of the question.
Practical consequence: "being cited by AI" is not a single objective but one per engine. Without multi-engine monitoring, you don't even know which version of your question the engines are answering.
Observation 3: small brands risk identity confusion
When asked directly about our brand, a conversational AI engine (detection from August 12, 2026) responded by citing an online gaming server, a Minecraft map, and a forum profile with names similar to ours. None of the cited sources were us. When asked about brand + product, the same type of engine correctly cited the founder's LinkedIn profile instead.
The lesson is the part of GEO that is talked about least: entity SEO. A young brand, without a consolidated presence on consistent external sources (LinkedIn, company profiles, third-party mentions), is just an ambiguous string to a language model. The Schema.org Organization markup with sameAs, name consistency everywhere, and external mentions are not technical details: they are what decides whether the AI's answer talks about you or about someone with the same name.
Observation 4: AI Overview sets prices (and follow-up questions)
For the sample service queries ("custom web development", "custom website development", Google searches from July 21, 2026) the AI Overview didn't just define the service: it reported explicit price ranges taken from the cited sources and closed the response with follow-up questions to the user ("What is the goal of your project? Do you have a reference budget?"), in a conversational format.
Two implications for service sellers: first, if the cited sources talk about prices and you don't, your market's price positioning is being written by others. Second, the AI response is already doing the lead qualification that your contact page used to do: getting inside that response is worth more than showing up just below it.
Observation 5: local ranking and neutral ranking are two different worlds
On the same query, on the same day (August 17–18, 2026), we observed our content in first position in the author's local SERP and outside the top 10 in the SERP queried from the datacenter. Those who assess their own visibility by looking only at their own browser — even in incognito mode — see a picture personalised by the IP geolocation.
For GEO this matters twice as much: the grounding of AI Overviews draws from the SERP the system sees, not the one you see. Reliable verification requires measurements from third-party IPs and, as a final reference, aggregated Search Console data — the only true average across all SERPs served.
What to do in practice with this data?
-
Bring key content into video too
Where consensus is built by Shorts and TikTok, a short video derived from a guide is a citation asset, not an extra. Text remains the foundation; video is the format that, for some queries, gets anchored to the answer.
-
Publish pillar content in English
Conversational engines build answers from pools of English-language sources even for Italian queries. The English version of your key content is the gateway to those answers.
-
Consolidate your entity before scaling content
Schema.org Organization with sameAs, a consistent name everywhere, active external profiles. If the model doesn't know who you are, it can't cite you — or worse, it cites a namesake instead.
-
Measure by engine, with historical data
Answers vary across engines and over time. Only saved, comparable snapshots distinguish a real pattern from a one-off observation — the same principle we applied when writing this article.
The techniques to put these priorities into practice are in the operational guide to AI Overview optimization (with the 35-point checklist) and in the article on how to prepare content for generative search engines. To understand the fundamental difference between optimizing ranking and optimizing citations, read GEO vs SEO: what is the difference.
Frequently asked questions about this data
Not necessarily: the sample consists of 7 vertical queries on software, SEO, and AI content, collected between July and August 2026. The patterns (weight of videos, anglophone sources, entity confusion) are engine mechanisms, so it is plausible to find them elsewhere, but the only way to know is to track the queries in your own sector. That is exactly why we publish the method alongside the data.
The AI citation tracker integrated into AraLith COS: for each tracked query, it periodically interrogates Google (SERP + AI Overview) and conversational AI engines, saving the full response, cited sources with snippets, the status of target content, and organic positions. Snapshots are stored in history and can be compared over time.
Because they combine three sources of variability: the underlying search index (which changes), the model's generation (which is non-deterministic), and the localisation of the request (IP, language). In our surveys, the same query produced different responses on the same day. This is why a single check proves nothing: you need the historical record.
No. In our sample we observed content in first position in the local SERP that was absent from both the neutral top 10 and the cited sources. Grounding starts from organic visibility but selects according to its own criteria: extractable format, source type (for some queries, video and social) and entity recognisability.
Do you want this data on the questions in your industry?
The AraLith COS AI tracker monitors Google's and AI engines' responses to the questions that matter for your business, with history and status of your content.
