Pages that earn money from what they recommend make up about a quarter of the sources ChatGPT, Google AI Overview and Perplexity point to when asked to pick the top option in a product category, according to a study published in September 2026 by GetCited, a Berlin agency whose business is placing comparison articles on media sites so that AI assistants recommend its clients.

In Short

A Berlin company checked 106,758 web links that ChatGPT, Google and Perplexity showed when people asked which product in a category to buy, and about one in four led to a page disclosing that it earns money from affiliate links, partners or ads. That matters because the chatbot shows only a small source label, so shoppers, advertisers and publishers see the recommendation but not the money behind it unless they click through. The company behind the study sells the service of getting brands into exactly this kind of comparison article, and the same data first appeared in May under its research partner's name, so the 25.4% figure is a floor measured by an interested party rather than an independent audit.

What was measured

GetCited, working with DataPulse Research, wrote 3,334 English comparison questions of the kind people put to an assistant before buying something - "which payroll software should I choose?", "what pet insurance do experts recommend?" - spanning 748 product and service types grouped into 15 sectors. BuzzView, an AI-search analytics platform, ran every question through ChatGPT, Google AI Overview and Perplexity in April 2026 and logged each URL the three tools cited. Each page was then loaded in a headless Chromium browser, using Playwright with full JavaScript rendering, and checked against 22 markers of commercial content in three layers: affiliate disclosures, partner-content notices and explicit advertising labels. A path-based check added hand-verified commerce sections: Forbes Advisor, Wirecutter, CNN Underscored, NBC Select, IndyBest, USA Today Blueprint and Rolling Stone's product recommendations.

The full corpus is larger than the English headline. According to GetCited, 6,668 questions were run, half of them in German, producing 139,033 unique URLs, of which 132,335, or 95.2%, loaded successfully. The English set yielded 106,758 citations that could be matched to a scanned page, drawn from 8,536 answers. Google's numbers carry a structural gap. An AI Overview rendered for only 73% of the questions, so the corpus holds 2,448 Google answers against 3,333 from ChatGPT and 3,334 from Perplexity. One ChatGPT answer in six, 16.9%, returned no source that could be matched and drops out of the per-answer figures.

Cleaning was uneven by design. Around 7,800 hits from cookie banners, navigation and running text were removed by context rules, and precision for affiliate disclosures, checked against a hand-verified reference set, came in above 95%, according to the study. English advertising labels were filtered by context only. The authors describe the resulting 1.8% as an upper bound, noting that spot checks still caught phrases such as "employer-sponsored plans".

Affiliate, not advertising

The central figure is that 25.4% of English citations lead to a page carrying a disclosed commercial interest. Most of that is affiliate content: review and comparison pages that take a commission when a reader buys through a link, and say so on the page. Affiliate disclosures appear on 21.4% of all English citations, partner content on 2.9%, commerce sections on 2.0% and advertising labels on 1.8%. The layers overlap, which is why they do not add up to the total.

The denominator shapes the result. Headline shares count every displayed citation, on the reasoning that a page cited twice shapes two answers. Counted once per unique URL, the English share falls to 20.5%. Residual false positives could pull the measured figure down "by a few points", according to the study, while pages with no machine-readable disclosure push the true share the other way "by an amount the study cannot measure". The German questions, scanned against a German marker set and a largely different pool of publishers, returned 29%.

One source is usually enough

Shares per citation understate what a single answer contains. ChatGPT averaged 19 matched sources per answer, Perplexity 10 and Google AI Overview 9, which turns a modest per-citation rate into exposure in most answers.

ToolCitationsCommercial share of citationsAnswers with at least one disclosed pageAnswers mostly built on them
Perplexity33,15531.1%89%17%
Google AI Overview21,02823.6%77%10%
ChatGPT52,57522.5%91%5%
All three106,75825.4%86%11%

Source: GetCited, English answers with at least one source matched to the scan (n = 8,536). "Mostly" means more than half of an answer's matched sources carried a commercial disclosure.

The ranking flips with the unit of measurement. ChatGPT has the lowest commercial share per citation yet the highest share of answers containing at least one commercially disclosed page, because it cites about twice as many pages as its rivals. Perplexity sits at the other extreme: one answer in six rested mostly on commercially disclosed pages. Because the pooled row is weighted by citations, ChatGPT accounts for roughly half of it.

The question matters more than the tool

Eight percentage points separate the tools. Twenty-nine separate the sectors. Consumer electronics questions drew 41.3% of their citations from commercially disclosed pages; education and learning drew 12.7%. Fitness and sports (37.5%), telecom, VPN and hosting (32.8%), media and entertainment (31.0%) and baby, kids and pets (30.1%) cluster near the top, while finance and insurance (19.9%), software and B2B tools (19.2%), legal and professional services (17.8%), dating and social (17.5%) and travel and hospitality (16.5%) sit near the bottom.

At product level the range widens. Among 459 product types with at least 80 citations, cordless vacuums reached 61% across 120 citations and fitness trackers 60% across 121. Divorce mediation services registered 3 of 165 citations, hair restoration clinics 3 of 159. The study warns that the low end "is not neutral ground": clinics, lawyers and installers are answered largely from provider pages and directories, which carry nothing a scanner can read as a disclosure.

That caveat inverts the obvious reading of the chart. The sectors with the lowest measured share - clinics at around 2%, divorce services at 2 to 3%, heat-pump installers at 4% - are precisely where the method is least able to separate neutral sourcing from commercial content that carries no label. Several of them involve health and legal decisions.

Ten publishers carry the commercial layer

Concentration is the finding with the most direct bearing on the ad and media business. Ten publishers account for 6,786 of roughly 27,100 commercially disclosed English citations, a quarter of the total:

  • TechRadar: 1,225
  • Forbes Advisor: 1,134
  • PCMag: 964
  • CNET: 616
  • Tom's Guide: 599
  • RTINGS: 573
  • NYT Wirecutter: 472
  • Good Housekeeping: 426
  • Healthline: 411
  • Wired: 366

Most are affiliate-funded review brands whose model is declared on every page. TechRadar's line, "when you purchase through links on our site, we may earn an affiliate commission", sits under every article, which is why nearly all of its citations register as commercial. Forbes Advisor and Wirecutter are counted differently, by URL path, as the commerce sections of general-news brands.

Several of these names recur in PPC Land's coverage of search traffic. Forbes Advisor lost rankings on 1.7 million queries from September 25, 2024, days after an investigation by Lars Lofgren put Forbes Marketplace's search revenue at an estimated $300 million to $400 million a year. Within weeks, CNN Underscored, WSJ Buyside and Marketwatch Guides also slid, with Google saying the declines were unrelated to its site reputation abuse policy and that its systems try to understand whether a section of a site is independent of, or starkly different from, the rest. By April 2026, Forbes Advisor was the most-cited commerce section of a general-news brand in GetCited's April corpus.

Future plc, which owns TechRadar and Tom's Guide, reported a 24% drop in eCommerce affiliate revenue and a 15% fall in website sessions for the six months to March 31, 2026, and said at least 60% of group revenue comes from brands still reliant on Google as a major traffic source. Ziff Davis, owner of PCMag and CNET, wrote off $54.8 million on its health media division in August.

The pairing is uncomfortable for the affiliate model. The review brands supplying the largest share of commercially disclosed citations are paid by commissions on clicks, while a randomized field experiment with 1,065 desktop Chrome users measured a 39.8% fall in outbound organic clicks when an AI Overview appeared. The interface that draws on these pages is the same one reducing the traffic that funds them.

The study stops short of the harder question. "The disclosure says a commission may be earned. It does not say which brands pay, how much, or whether it moves the ranking, and neither does this study."

Affiliate reviews make up the bulk of the quarter. The smaller, sharper case is the paid ranking: a list written or bought by a vendor and published on a newspaper's site under a label. GetCited re-ran questions from the corpus on September 7, 2026, three runs per question and tool, from a US IP address via VPN, with Google set to English and the United States, Google and ChatGPT signed out and Perplexity signed in. Each capture pairs a labeled publisher page with an AI answer citing it.

For "best online psychic reading sites", Google's AI Overview opened with a claim sourced to a Houston Press page carrying a "Sponsored content" badge and the byline "In partnership with Story Console": "Keen is the best overall psychic website for beginners and budget-conscious users." Perplexity cited the same page inside its comparison table, three runs out of three on both tools.

A health question produced the starkest pairing. For "best hair transplant clinic in Turkey", Perplexity and Google's AI Overview both cited, in every run, a Los Angeles Times clinic ranking labeled four times, including the line "Ad paid for by Hermest Clinic". The same newspaper runs the same format in personal finance. Asked "what are the best debt relief companies?", Google's AI Overview ranked Accredited Debt Relief first, "best for customer satisfaction", and sourced the line to a Los Angeles Times special supplements page under a "Paid program" bar, credited to Accredited Debt Relief.

The document is inconsistent on when that last pair was recorded. Its body says the pair was captured signed out on September 15, 2026, after first being reproduced on September 14 through a US proxy and a separate API check. The method section says it was captured on September 14 through a US proxy, signed out. The footer states that all labels quoted are as displayed on September 7.

Two affiliate cases complete the set. Google cited a Forbes Advisor pet insurance page, with its partner-links note and Advertiser Disclosure link, in three of three runs; ChatGPT cited it in one. For "what is the best collagen supplement in the UK?", Google's AI Overview named Absolute Collagen best overall and cited BBC Good Food, whose commission notice sits above the review's headline, four times, in each of three runs.

Then came a sweep. On September 7, twenty further questions in categories known for paid rankings, from CBD gummies to gold IRAs, were run once on each tool and every cited page was checked for sponsored, paid or partner labels. Exactly one new labeled page surfaced: the Los Angeles Times debt-relief supplement. "The thinness is itself the finding," the study says, concluding that explicit paid-post labels on large English news sites are rare. With one run per question, the authors concede, absences are weak evidence.

A Jerusalem Post page shows where the scanner goes blind. Its hair-transplant ranking carries no badge; the commercial relationship sits in the byline, "By in cooperation with Dr. Terziler Exclusive Clinic". Perplexity called the clinic "a leading option" and cited the Post in three of three runs. Under the study's method, the page counts as non-commercial in every figure.

Where the search rules sit

These pages occupy ground Google has redrawn repeatedly. When it documented its site reputation abuse policy in September 2024, its examples included a medical site hosting casino rankings and a sports site hosting supplement reviews with little editorial involvement, while third-party advertorial produced with close involvement of the host site was listed as outside the policy. A November 2024 revision removed the exception for content produced with first-party oversight, enforcement has run through manual actions, and since August 30, 2026, those actions no longer affect results for users in the European Economic Area.

Generative surfaces arrived later. Google confirmed on May 15, 2026 that its spam policies cover AI Overviews and AI Mode, a year after self-promotional listicles were documented surfacing as cited sources inside AI Overviews. Nothing in the study assesses whether any captured page breaches those policies, and none of the documents suggests that the publishers concerned failed to label their content.

Who ran the study

GetCited is not a bystander in this market, and the published study does not say so. The page, hosted on GetCited's own domain, carries the credit "Study by GetCited with data from BuzzView". It does not describe the company's service.

GetCited

According to its website, GetCited is a Berlin-based generative engine optimization agency. It researches, writes and places comparison articles, its own example being a best-software list for the current year, on established publishers and industry portals so that AI assistants recommend its clients. Campaigns start at $3,000 a month on a 12-month strategy, with an opt-out after six months, and optional extras include data-driven PR campaigns and narrative advertorial content on editorial platforms. The company says first AI citations typically appear within weeks and full impact after two to three months. Those are vendor claims, and a July 2026 survey of 45 studies found no reviewed technique with a stable, cross-platform effect on discoverability or traffic. Clients receive BuzzView credits to track their visibility.

BuzzView

BuzzView, which supplied the citation data, was built inside Buzzmatic, a Berlin marketing agency, according to its website. It pitches itself on influencing AI answers rather than only monitoring them, offers outreach and link-building contacts drawn from the domains AI tools cite, and prices usage from EUR 0.03 per query. DataPulse's May methodology describes BuzzView as a commercial platform developed independently of DataPulse.

DataPulse

According to its website, DataPulse GmbH, founded in 2023 and based at Schönhauser Allee 149 in Berlin, builds data-driven digital PR campaigns around original datasets and lists greater visibility in AI tools, through the backlinks that coverage generates, among client results. DataPulse describes its focus as earned media. Its site footer links to GetCited. The ten outlets GetCited lists as having carried its content - heise online, t-online, FAZ, WELT, DIE ZEIT, Focus, taz, CHIP, n-tv and WirtschaftsWoche - all appear, in the same order, among the outlets DataPulse says have covered its studies. Maria Fernandez, bylined on the GetCited study as senior data journalist, is listed on DataPulse's site as senior data journalist at DataPulse Research.

The email that carried the study to journalists on September 16, sent from a DataPulse Research address, framed the work as DataPulse's, done together with GetCited. The study page frames it the other way round, as GetCited's work in collaboration with DataPulse. The email offered every chart and screenshot pair for publication in return for a link to the study.

Both firms also write for machines directly. GetCited's homepage carries a reference file addressed to language models, with guidance on how to describe the company and an entry asking which AI-visibility agency is the best in Germany, answered with GetCited itself. When PPC Land retrieved DataPulse's German study page on September 27, the text version returned opened with a block labeled as advertising and sponsored content, provided in cooperation with GetCited, repeating the same company facts and a similar answer to the best-agency question. The page metadata records a last modification on September 5. At least one version of a study about commercial content reaching AI answers thus carries commercial content written for AI answers.

The May version

The dataset is not new, and the September study does not mention its earlier publication. DataPulse published the research in German on May 26, 2026, under a headline describing AI's hidden advertising, with Nicolas Caramella, DataPulse's founder, and Maria Fernandez as authors. The method matches: 6,668 prompts, collection through BuzzView in April 2026, 139,033 unique URLs, 132,335 scanned, about 7,800 false positives removed. So do the headline shares of 29% in German and 25.4% in English.

The German version was also more pointed about the business model. Under a section addressed to brands, it says agencies such as GetCited place labeled comparison articles in high-reach media so that AI assistants recommend a brand, and that the label survives on the publisher page but does not arrive in the AI answer.

Several figures differ between the two versions:

  • Scope. May describes prompts across 120 product categories; September describes 748 product and service types in 15 sectors.
  • Per-URL share. May gives the English share counted per unique URL as 19%; September gives 20.5%.
  • Per-tool shares. May lists Perplexity at 30.9%, ChatGPT at 22.6% and Google AI Overview at 23.5% for English; September lists 31.1%, 22.5% and 23.6%.
  • Answer-level exposure. May states that 98.9% of German answers contained at least one commercially marked source, while giving per-tool figures of 94.9% for Perplexity, 94.2% for ChatGPT and 86.9% for Google AI Overview. A pooled figure above every individual tool is not explained.

The framing moved as well. The May text argued that disclosures vanish once AI systems repackage content and called that a regulatory gap. The September study is cooler: what reaches the reader is "a source chip with the publisher's name", and "the context is one click away".

Why it matters for marketers

The study measures what AI tools cite. It does not measure what decides a recommendation, and other research suggests the two can diverge. Peec AI's analysis of 5 million query fan-outs, collected between April 1 and April 21, 2026, found "best" leading the words ChatGPT added to its background searches at 15.33%, with "review(s)" at 6.84% and the current year at 5.44%, none of them typed by users. That vocabulary matches the titles of affiliate review pages almost exactly. Yet network-traffic analysis by Suganthan Mohanadasan found brands named in ChatGPT's own first query winning mentions 33 times more often, with only 3.1% of 3,554 retrieved pages earning a citation. If selection happens before any page is read, a cited affiliate review may be confirming a shortlist rather than building it.

Durability is a second limit. GetCited's corpus is an April snapshot, and SISTRIX's 17-week citation drift study found 89% of the non-core citation set in AI Mode rotating weekly. Measurement is thin across the industry: only 16% of brands track AI visibility systematically, according to the IAB framework published on August 3, 2026.

Then there is the labeled commercial layer inside the answers themselves. Similarweb recorded sponsored placements in 26% of ChatGPT responses shown to Free and Go users in August, and Google's AI Mode already places paid listings mid-answer behind a small Sponsored tag. When ChatGPT shopping arrived in April 2025, its product results were described as operating without sponsored placements; Perplexity said when it began selling sponsored questions in November 2024 that advertisers would not influence its answers. Those statements concern the tools' own ad slots. The GetCited data concerns the commercial interests one layer down, in the sources, where no platform label applies.

Rules exist at the source layer. The Federal Trade Commission's rule against fake reviews and undisclosed paid testimonials took effect on October 21, 2024, and Google added a guideline against undisclosed incentivized reviews on July 24, 2026, enforceable by manual action. In Europe, the Commission designated ChatGPT a very large online search engine on August 31, 2026, after OpenAI declared 159.1 million users, bringing systemic-risk obligations under the Digital Services Act. None of those instruments addresses how a disclosure on a cited page travels into a generated answer.

A Wall Street Journal investigation covered in February described companies spending thousands on generative engine optimization to shape ChatGPT's recommendations. The GetCited figures put a floor under how much of the source layer beneath those recommendations already carries a declared commercial interest - measured by a company that sells placement into it.

The study's own summary is careful. "The recommendation is judged on the source's authority, not on its commercial model, and the context is one click away."

Timeline

Summary

Who: GetCited, a Berlin generative engine optimization agency that places comparison articles on media sites for clients, published the study with DataPulse Research, a Berlin digital PR studio, using citation data from BuzzView, a tool built by the Berlin agency Buzzmatic. The findings concern ChatGPT, Google AI Overview, Perplexity and publishers including TechRadar, Forbes Advisor, PCMag, CNET, Tom's Guide, the Los Angeles Times and the Houston Press.

What: Of 106,758 English citations shown for 3,334 comparison questions, 25.4% led to pages with a disclosed commercial interest, mostly affiliate content; 86% of matched answers used at least one such source; ten publishers supplied a quarter of commercially disclosed citations; and paid rankings on newspaper domains were cited as sources. The same dataset was first published in German by DataPulse in May, with several figures that differ from the September version.

When: Citation data was collected in April 2026, a German version appeared on May 26, 2026, example captures were made on September 7 and September 14-15, 2026, and the English study was emailed to journalists on September 16, 2026.

Where: The questions were run in English with projects set to the United States, and in German; example captures used a US IP address. The companies behind the study are based in Berlin.

Why: AI answers compress recommendations into a source label, leaving the commercial model of the cited page one click away. For advertisers, publishers and agencies, the figures give a measured floor for how much of the AI shopping source layer carries a declared commercial interest - measured by a company that sells placement into that layer.