Reddit dominates Google’s AI Overviews. The assistants barely cite it.

The counts are small on both sides. Reddit was 13 citations in 166 for AI Overviews on shopping questions (a 95% Wilson interval of 4.6% to 12.9%), and 15 in 1,378 for the assistants. The gap between the surfaces is larger than that uncertainty. What this study cannot say is why: it measures what each surface cites, not how either chooses.

The five assistants do not behave like one system. Perplexity returned 44,256 of the 151,559 citations, 29.2% of the pool, from the same 5040 answers that gave ChatGPT only 13,150. A pooled percentage therefore mostly describes Perplexity. That is why every figure that matters here is also shown per assistant, and why the balanced Reddit shares above average each assistant’s own share instead of counting citations. Balanced this way, Reddit was 0.8% of shopping citations against 1.1% pooled, and 0.8% of corporate citations against 0.9% pooled.

The most-cited domain differs by assistant, so a domain that looks small in the pooled list can be a top source for one assistant. With 5040 answers each, one answer is worth about 0 percentage points of an assistant’s share of answers, so small gaps between assistants are not evidence of a preference. The gap between Perplexity and ChatGPT in citation volume is large enough to matter.

Only 5 of the 10 most-cited domains for shopping questions also appear in the top 10 for corporate reputation questions (forbes.com, nerdwallet.com, consumerreports.org, reddit.com, businessinsider.com). The shopping list adds nytimes.com, pcmag.com, cnet.com, tomsguide.com and rtings.com, mostly review, ratings and product-testing publishers. The corporate list leans on ratings bodies, business and finance press and sustainability outlets.

Reddit sits in both lists at a similar level: 1.1% of shopping citations and 0.9% of corporate ones. It is one source among many for the assistants, and no domain exceeds 1.1% of the pooled citations, so the citations spread across a long tail of 34,802 domains.

For the same 105 questions we also recorded the AI Overview Google returned, on September 28, 2026. Of 105 questions, 93 triggered an overview (32 of 40 shopping questions and 61 of 65 corporate ones). Those overviews made 447 citations to 312 unique domains. Of the ten most-cited domains in AI Overviews, 7 are social or user-generated platforms (youtube.com, reddit.com, facebook.com, quora.com, instagram.com, linkedin.com and tiktok.com). Of the ten most-cited domains in the assistants, 2 are such a platform (reddit.com and youtube.com).

Community sites (Reddit, Quora, Stack Exchange, Hacker News, Trustpilot, Glassdoor) were 8.7% of AI Overview citations and appeared in 38.7% of overviews. For the assistants they were 1.7% of citations and appeared in 8.4% of answers. Facebook, YouTube, Instagram, TikTok and LinkedIn are outside that community set, and several of them rank above it in AI Overviews.

This is the pattern behind the popular claim. A person who searches Google for a product recommendation sees an overview that cites Reddit far more often than the assistants do. A person who asks an assistant the same question gets an answer built mostly from review publishers. Both are “AI answers”, and a study of one says little about the other.

Counting citations to newswire websites measures the wire’s own domain, not its reach. A release is published once and republished by finance portals, local news sites and trade aggregators, and the citation then carries the republisher’s domain. We therefore looked for wire content at three confidence levels and did not add them together.

The direct count is 475 citations (0.3% of the pool), spread across 17 wire domains (prnewswire.com 219, businesswire.com 132, globenewswire.com 52, openpr.com 21, einpresswire.com 16, newswire.com 10, prweb.com 10, edge.prnewswire.com 3, secure.businesswire.com 3, mma.prnewswire.com 2, newsfilecorp.com 1, pressat.co.uk 1, issuewire.com 1, abnewswire.com 1, prnewswire.co.uk 1, 24-7pressrelease.com 1, cts.businesswire.com 1). Perplexity contributed the most, 171. The URL-name signal found 47. The path-hint signal found 2371, but it is not counted as wire: jdpower.com alone accounts for 635, and organizations use these paths for their own announcements.

The page-body pass covers what the URL cannot show. For 125,896 unique cited URLs with no other wire signal, it fetched the page and looked for a wire dateline, the attribution line syndicated copy keeps verbatim, such as “/PRNewswire/ --” or “(BUSINESS WIRE)--”. Of 96,386 pages it could open, 79 carried one. 29,510 of 125,896 URLs (23.4%) could not be opened, usually because the site rejects automated requests, and those are unknown rather than negative.

A different signal turned up as well. 729 of the pages that opened are articles that hyperlink a release on a wire service as a source, such as a newsroom post or a trade-press story citing a PR Newswire release. That is not syndicated copy, and it is counted separately. It is the pattern that shows up in this data: a release informs coverage, and the assistant cites the coverage.

How the method changed, since it moves the headline figure. A first version counted a page whenever a wire service’s name appeared in its HTML. Checking those matches by hand showed they were a script listing wire-feed providers, hyperlinks to releases, and similar page furniture rather than syndicated text. Requiring a dateline removed all of them. 373 opened pages still name a wire service in their visible text without a dateline; they are not counted.

One known blind spot. Some publishers load a syndicated release into the page with JavaScript, so its dateline is not in the HTML this check reads. The Globe and Mail’s press-release section works this way; those URLs are caught only by the path signal above, which is why path hints are reported rather than discarded.

The finding: in this sample, wire services were a small direct source of AI citations (0.3%). None of the shopping-question citations were wire domains. 79 reachable pages carried syndicated wire copy. Where releases do appear, it is mostly second-hand, through articles that cite them. The unreachable and JavaScript-rendered pages mean the true figures could be higher, so treat every wire number here as a floor. What the study cannot show is whether distributing a release caused any citation. That needs a counterfactual, such as the same release with and without distribution, which this snapshot does not have. It measures presence, not effect.

The sector figures cover the 25000 answers to corporate reputation questions (5 assistants, 250 questions per sector). Consumer questions are grouped by product category rather than sector and are not in this table. Each sector has 1250 answers, so the rates are coarse.

We classified each sector’s top-cited domains on corporate questions by source type (owned, news, trade, review and affiliate, community, video and social, other). Cells show the share of that sector’s classified citations, with the count in brackets. Across all sectors the largest types were review & affiliate (36.4%), news (16.9%) and trade (15.8%).

Owned sources. Domains belonging to one of the measured brands totalled 1769 of 12842 classified citations (13.8%): airbus.com (Aerospace & Defense, 68); lockheedmartin.com (Aerospace & Defense, 66); openai.com (AI Companies, 142); anthropic.com (AI Companies, 101); help.openai.com (AI Companies, 70); sephora.com (Beauty & Personal Care, 73); sentinelone.com (Cybersecurity, 86); crowdstrike.com (Cybersecurity, 85); fortinet.com (Cybersecurity, 69); chime.com (Fintech & Payments, 96); paypal.com (Fintech & Payments, 78); hilton.com (Hotels & Travel, 62); fedex.com (Logistics & Delivery, 145); uber.com (Ride-Hailing & Delivery, 242); lyft.com (Ride-Hailing & Delivery, 101); help.uber.com (Ride-Hailing & Delivery, 84); help.lyft.com (Ride-Hailing & Delivery, 60); support.google.com (Technology, 47); verizon.com (Telecom & Media, 94). A measured brand’s own site was not among the most-cited sources in 10 of the 20 sectors. An automated “owned” label only counts when the domain matches a measured brand.

Limits. The classifier ran only over each sector’s eight most-cited domains, so this describes roughly the top 8.6% of corporate citations, not the tail, where a brand’s own pages may appear once each. Its labels are automated and unaudited. And “rarely cited” describes these questions, which compare companies; a question naming one company may cite that company’s own pages more often.

1. Corporate questions. Corporate and reputational category questions per sector (e.g. 'which US airlines are the most reliable?'). Five unbranded questions per sector across 20 sectors (5000 questions), each sent once to each assistant: 25000 answers.
2. Consumer questions. Unbranded consumer shopping questions across ten categories (e.g. 'what is the most reliable used car under $20,000?'). 40 questions across 10 categories, unbranded, phrased the way a person types into an assistant and spread across categories so no single vertical dominates. Each was sent once to each assistant: 200 answers. Both question lists are published below.
3. Assistants. The five assistants the PeakMetrics product benchmarks, each on the model the product uses, with web search enabled, called through an API: Perplexity (sonar), Gemini (google/gemini-3.1-flash-lite), ChatGPT (gpt-5.4-mini), Claude (claude-haiku-4-5), Grok (x-ai/grok-4.3). Consumer apps can run different models and search settings, so these answers may differ from what the apps return.
4. Google AI Overviews. The same 105 questions were run through Google Search and the AI Overview recorded (source: DataForSEO SERP google/organic/live/advanced, US English, desktop; measured September 28, 2026). Not every query triggers an overview: 93 of 105 did (32 of 40 shopping questions, 61 of 65 corporate). Only triggered overviews enter the AI Overview figures, and they are reported separately from the assistants, never pooled.
5. Citations. Every URL in an answer’s annotations, or in an overview’s references, is recorded. The headline unit is one citation per distinct domain per answer, so a domain cited three times in one answer counts once. The wire body pass uses unique URLs instead.
6. Balanced shares. A provider-balanced share is the mean of each assistant’s own share, so an assistant that cites many sources per answer does not outvote one that cites few. The pooled share is shown alongside.
7. Wire detection. Three signals: the hostname matches a list of wire services (PR Newswire, Business Wire, GlobeNewswire, PRWeb, EIN Presswire, ACCESSWIRE, Newsfile and others); a wire service’s name appears in the decoded URL; the URL contains a release-style path. The body pass then fetches each remaining URL (identifying itself as a research bot), reads the first 120 KB and looks for wire names and datelines. Any fetch failure is recorded as unknown.
8. Community. Reddit is matched by hostname. The wider community set is Reddit, Quora, Stack Exchange, Stack Overflow, Hacker News, Trustpilot and Glassdoor. Other social platforms, such as YouTube and Facebook, are counted by domain in the top-domain lists and are not part of the community set.
9. Source type. Each sector’s eight most-cited domains on corporate questions were classified by an LLM-based classifier with a 0.6 minimum confidence; lower-confidence labels were dropped.