Best-product answers run on affiliate reviews

A best-product answer is an AI engine's summary of other people's product comparisons, and on commercial questions a large share of those comparisons are written by publishers who earn a commission when you buy.

That is the plain reading of a study published in September by GetCited and DataPulse Research. They asked ChatGPT, Google AI Overviews and Perplexity comparison questions ("best X compared", "which X should I choose") and then checked every cited page for a commercial disclosure. The result changes how a brand should think about getting recommended, and how an AI visibility audit should classify the sources it finds.

#What the study measured

The GetCited study generated 6,668 comparison questions (3,334 in English) across 748 product and service types, set the engines to English and the United States, and collected 106,758 citations in April 2026. A headless browser loaded each cited URL and looked for machine-readable markers: affiliate disclosures, partner-content notices, advertising labels and commerce sections identified by URL path.

The headline result: 25.4% of citations disclosed a commercial interest. Affiliate disclosures made up most of that, at 21.4% of citations. Explicit sponsored labels were rare, under 2%, and the authors flag their 1.8% advertising-label figure as an upper bound because those labels were filtered by context only.

The authors call 25.4% a conservative reading. A page with a commercial relationship but no machine-readable disclosure counts as non-commercial in their data, and 4.8% of cited URLs could not be loaded at all. Treat the number as a floor.

#Almost every answer leans on at least one

The per-citation share understates exposure, because engines cite many pages per answer. The study reports this per engine:

Engine	Citations	Commercial share	Answers with at least one commercial source
Perplexity	33,155	31.1%	89%
Google AI Overview	21,028	23.6%	77%
ChatGPT	52,575	22.5%	91%

ChatGPT has the lowest per-citation share and the highest per-answer exposure. It cites about 19 pages per answer against about 10 for Perplexity, according to the same study, so even a modest share almost guarantees that some affiliate page shaped the answer. One caveat applies to the ChatGPT row: 16.9% of its answers had no sources the researchers could match.

#Category beats engine

The useful finding for planning is the spread. In the GetCited data, commercial sourcing ran from 41.3% in consumer electronics, 37.5% in fitness and sports and 32.8% in telecom, VPN and hosting, down to 16.5% in travel and hospitality and 12.7% in education. The authors put it directly: the gap between engines is eight points, the gap between categories is 29.

That means "which engine should we optimize for" is the second question. The first is what kind of source your category's answers are built from. A laptop brand and an online course provider are playing different games even when they track the same three engines.

#Ten publishers carry a quarter of it

Concentration is high. Ten publishers supplied about 25% of all commercially disclosed citations in the study. The top three were TechRadar (1,225 citations), Forbes Advisor (1,134) and PCMag (964), followed by CNET, Tom's Guide, RTINGS, Wirecutter, Good Housekeeping, Healthline and Wired.

For a brand in a high-commercial category, this list is closer to a media plan than a curiosity. If your product is absent from the comparison pages these sites publish, it is likely absent from the pool the engine summarizes, no matter how good your own product page is. We made a related point about listicles that recommend your rivals: the page that gets cited on a comparison question is usually not the vendor's own.

#Labels did not keep pages out

The study also documents labelled commercial pages that engines cited anyway. A Google AI Overview opened an answer about psychic readings with a claim sourced to a Houston Press page labelled sponsored content. Perplexity cited a Los Angeles Times page labelled as a paid program for a hair transplant clinic. A Google AI Overview called a debt relief company "best for customer satisfaction", sourcing the claim to a Los Angeles Times page credited to that company.

Google's own spam policies cover parts of this ground. Site reputation abuse applies when third-party content sits on a host site "mainly because of that host's already-established ranking signals", and the same document lists attempts to manipulate generative AI responses in Search as spam. Those policies are enforced through ranking and manual actions. Nothing in the study suggests the answer layer applies a separate check for disclosures on the page it quotes.

#Disclosure is written for people, not parsers

The FTC's endorsement guidance says "the closer the disclosure is to your recommendation, the better" and warns that a disclosure separated from the review may be missed. That standard is about a reader seeing the review and the disclosure together.

An AI answer breaks that pairing. The engine extracts a recommendation, cites the page and drops the disclosure. The user who reads "best for customer satisfaction" in an AI Overview never sees the label on the source page unless they click through. Disclosure rules were designed for the page, and the answer is not the page.

#What to do with this

Classify your cited sources before you benchmark. An AI visibility report that lists citing domains without tagging which ones are affiliate, sponsored or editorial hides the most important variable in commercial categories. Tag the citing pages by commercial status and report the share per category, the same way the study did.

Find out what your category is built on. Run your own comparison prompts and count. If a third or more of citations come from affiliate reviewers, your path into the answer runs through those reviewers' comparison pages, and product sampling, review programs and accurate spec sheets are the work.

Give reviewers something real to test. Google's guidance on high quality reviews asks reviewers to show evidence of their own experience and to explain what sets a product apart from competitors. Publishers that follow it need hands-on access, measured performance and clear differentiation. A brand that supplies those gets better coverage than one that supplies a press release.

Publish the comparison facts yourself. Affiliate pages win partly because they compare. A vendor page that states specs, limits, pricing and which use case the product fits, in plain text, gives the engine a first-party source to check claims against.

Read paid placements as a risk. The study shows labelled paid pages can reach AI answers. It also shows those placements sit on news domains that Google's site reputation policy addresses. Buying a "best of" slot on a host site is a bet on a page that can lose its rankings, and with them its retrieval, after a policy change.

Keep the floor in mind. Every number above counts only machine-readable disclosures. Your own audit will undercount for the same reason, so treat any commercial share you measure as a minimum.