Companies that publish best-of lists placing themselves first are recommended by AI assistants slightly more often when those lists are cited, according to research Scrunch published on September 24, 2026, while the effect on whether a brand is named at all runs several times larger.

In Short

Some companies write lists of the best tools in their market and put their own product in first place, and a new study checked whether AI chatbots then recommend those companies more often. That matters to any brand trying to appear in ChatGPT, Gemini, Claude or Google's AI answers, because those answers increasingly shape what people look up and buy. The effect is small: when your list gets cited, your odds of being recommended move from about four in 100 to about seven in 100, your odds of simply being named roughly double, and in one scenario the rivals on your list may benefit as well.

Nearly a third of cited comparison pages crown their author

Scrunch, an AI search visibility firm that describes itself as a Sitecore company, analysed roughly 10,000 URLs cited by AI assistants in responses collected between May 17 and June 30, 2026. The study covered seven answer surfaces: ChatGPT, Claude, Gemini, Microsoft Copilot, Perplexity, Google AI Overviews and Google AI Mode. Grok and Meta AI were not part of the dataset.

Of the nearly 8,000 pages the company could classify with its content taxonomy, comparison and evaluation pages accounted for 33%. Roughly 30% of those were self-promoting listicles, which Scrunch defines as content in which a company ranks its own product first in a "best [category]" list. Simple arithmetic puts that group at roughly 800 pages, close to the 818 treatment pages described in the methodology note at the foot of the post. Against them the company set a control group of 1,033 neutral third-party comparisons, pages whose publisher sells nothing on the list.

A response-text classifier then produced 29,382 labels, sorting each AI answer into recommend, neutral or caution. Scrunch hand-audited samples of both groups: 29 of 30 treatment pages were classified correctly, or 96.7%, as were 10 of 10 control pages. The post does not explain how the audited pages were chosen.

Authorship is not entirely clear. The post, filed under Scrunch's Data Studies category, carries the byline of Niharika Sharma. Material Scrunch circulated to press on the same day credits the study to Michael Ianelli, head of data science at Scrunch, who holds a Ph.D. in theoretical cloud computing. Neither document explains the difference. One plausible reading is that one person wrote the post and the other ran the analysis, but neither source says so.

A modest lift, and a test for momentum

The core comparison held three variables fixed - the prompt, the brand and the AI assistant - and compared answers that cited a brand's listicle with answers to the same prompt that did not. For the author brand, Scrunch's term for the company that published the list, the recommendation rate moved from about 4% to roughly 7% across AI assistants when its self-promoting listicle was cited, according to Scrunch.

That invites an obvious objection. Brands climbing in AI search for unrelated reasons may be both more likely to be cited and more likely to be recommended, which would produce the same pattern without the list doing any work. To check, Scrunch ran what it calls a falsification test. It first removed every answer that cited the listicle, leaving only responses the page could not have touched. It then split the remainder according to whether that same listicle was cited in a later answer to the same question. A citation that has not happened yet cannot shape an earlier answer, so any gap between those two groups measures momentum rather than influence.

The placebo gap came out at about a quarter to a third of the real one. Part of the observed lift therefore reflects brands that were already trending upward, on Scrunch's reading, but most of it does not. Applied to the rounded headline figures, that would leave an effect in the region of two percentage points attributable to the citation itself - an inference from the published ranges rather than a number Scrunch reports.

"When cited, the lift from self-promoting listicles is real, if modest," the company wrote.

Two further details complicate the picture. Pages listing four to 10 brands saw their author recommended more often than pages listing only two or three, or packing in 11 to 15, according to Scrunch, although the post gives no rates for those bands. And in 63.9% of responses citing a listicle, the assistant recommended nobody at all. It used the listicle as raw material for a balanced overview rather than adopting the page's first-place ranking as its own advice.

Mentions move further than recommendations

The larger effect sits in a softer metric. Scrunch separates a mention, where a brand is simply named in an AI answer, from a recommendation, where the assistant actively suggests it. AI answers that cited a brand's listicle named that brand about 39% of the time, according to Scrunch, against roughly 19% when the listicle was not cited.

Scrunch describes the mention effect as roughly four to six times the recommendation lift. The rounded figures - about 20 points against about three - would put the ratio closer to seven, which suggests the unrounded values behind the claim differ from those presented; the post does not publish them. The mention figure also carries a qualification the recommendation figure does not. "The mention lift hasn't been through this check yet," the methodology note states, referring to the falsification test.

Why does the distinction matter? An earlier Scrunch data study, cited in the post, found that when a brand is recommended rather than merely mentioned, users are nearly twice as likely to search for it on Google, visit its website and view its products on a retailer's product page. Research from other firms points the same way. Similarweb clickstream research published in June found AI-recommended brands 2.5 times more likely to receive a site visit within seven days, with 56% of that traffic arriving through branded search. The IAB's AI visibility measurement framework, published in August, treats Mention Rate as the most basic visibility metric: responses containing a brand divided by total responses.

The assistant sets the ceiling

Citation behaviour varies sharply by surface. When citing a neutral comparison page - an independent roundup in which the publisher sells nothing on the list - Gemini recommended a brand in 64.1% of responses and Claude in 49.9%, according to Scrunch. ChatGPT sat in the middle at 33.9%. Google AI Overviews did so in 28.7% of cases, Google AI Mode in 25.8%, Microsoft Copilot in 18.1% and Perplexity in 14.1%. Those four, in Scrunch's description, mostly summarise what they read.

The order shifts when the question narrows to how often the author of a self-promoting listicle is itself recommended. Claude leads at 15.6%, ahead of Gemini at 14.8%. Google AI Overviews follows at 8.9%, edging ChatGPT at 8.6%, with Google AI Mode at 7.7%, Perplexity at 6.3% and Copilot last at 4.8%. The gap between the top and bottom of that list is more than threefold. These rates cover every answer that cited the listicle, and Scrunch cautions that they are not directly comparable with the 4%-to-7% comparison, which was built to isolate the listicle's own effect.

The pooled figures blend two different kinds of product. Google's AI Overviews and AI Mode made up 31% of the treatment sample, 1,692 of 5,434 page-response pairs. That pair is Scrunch's working unit, since one response can cite more than one page; the 5,434 pairs came from 4,594 unique responses. On the two Google surfaces, the author brand was recommended 8.6% of the time, or 145 of 1,692 pairs. On the chat assistants the rate was 12.2%, or 455 of 3,742. The pooled figure of 11.0% sits between them.

Nor is the time window uniform. For six of the seven surfaces, Scrunch sampled the five most recent citing responses per page, which in practice covers June 10 to 30. Claude's data spans the full period from May 17. Noting that AI search systems change quickly, the company asked readers to "read every number as a snapshot of that window, not a constant."

The platform split sits awkwardly against earlier practitioner observation. SEO consultant Lily Ray has described reasoning models skipping low-trust listicle sources in their visible chains of thought, including the model behind ChatGPT's thinking mode. Scrunch's figures, collected over a few weeks in mid-2026, show Claude and Gemini as the assistants most willing to hand a recommendation to a list's author. The two are not necessarily in conflict. Ray's evidence consists of observations rather than measurements, and Scrunch's numbers describe what happens once a list has been cited, not how often lists are cited or passed over in the first place.

Naming rivals carries a conditional risk

AI assistants rarely lean on a single source. Across all answers that cited a self-promoting listicle, 43.4% also cited a competitor's own website, according to Scrunch, a pattern that correlated with a slightly lower recommendation rate for the author brand. Support in the other direction was almost absent: a page other than the author's own listicle backed the author brand in just 1.4% of answers. In most cases, then, the cited list was the only source vouching for its publisher, while the rivals named on it could draw on their own sites and independent coverage.

The most damaging outcome would be a ghost citation, in which a brand's content is cited but the brand itself is never named. Scrunch found that rare. The softer version was common. When a listicle was cited, a competitor was recommended instead of the author 24.3% of the time, a result the company says held under every cut it tried, including a platform-matched replication.

Does the list cause that? The raw numbers suggested it might. When a brand's own listicle was cited, the odds that at least one rival named on it was recommended ran 2.8 to 3.4 percentage points higher. Running the same falsification test on rivals, however, produced a placebo gap of 4.4 percentage points, larger than the effect it was meant to rule out. On that evidence, the rivals were most likely already rising for unrelated reasons, and the listicle simply happened to be cited around the same time.

Scrunch then split author brands and competitors into high- and low-visibility tiers, based on how often each is named as a rival elsewhere and how often it appears organically, and repeated the test across four combinations. When a high-visibility brand names high-visibility rivals, the author still registers its own lift and the rival effect fails the falsification test cleanly. When a lower-visibility brand names high-visibility rivals, the rival effect fails again, but the author's own lift is too small to measure. A high-visibility brand naming lower-visibility rivals produces a placebo that reproduces about half of the rival's lift, and Scrunch declines to call that result either way.

Only one combination survives. When a lower-visibility brand names lower-visibility rivals, the rival effect holds: a lift of 7 percentage points against a placebo of minus 1.9.

The company attaches its own caveats to that number. It is one of four scenarios, and no correction was applied for making four comparisons. A much thinner slice of the same data put the effect at 11 points, with too few observations to distinguish it from zero. Both lower-visibility scenarios already use every observation Scrunch holds, and the company says additional data will not sharpen them. The study also left one mechanism untested: whether linking to a rival from the page, as opposed to merely naming it, changes what gets retrieved. That is a live question for assistants built on retrieval-augmented generation, which fetch documents before composing an answer, and Scrunch treats it as open.

What the design can and cannot support

Scrunch divides its figures into two tiers. The descriptive numbers - presence, recommendation rates, co-citation and platform splits - come from the full run described above. The causal numbers, including the author lift and the rival tests, come from the tighter same-prompt, same-platform, same-brand comparison. That comparison was run on a capped sample covering roughly 6% of the eligible population. A full-scale run is planned to confirm the headline magnitudes, according to Scrunch, though the post gives no date.

The variable that differs between the two groups is whether an assistant happened to cite the listicle, not what is written on the page. The figures therefore describe what a citation does for brands already listed; they say nothing about what adding or removing a name would do. A looser version of the test, comparing a vendor's rate on its own listicle with its appearances elsewhere while holding the platform fixed, found no effect at all: a median of 0.0 percentage points across 61 cells, with a p-value of 0.62.

The prompts are Scrunch's own monitored questions. The study describes how AI assistants treat these pages in response to those prompts, not what buyers subsequently do. And the research comes from a vendor. Scrunch sells AI visibility monitoring, its post closes with an offer of a seven-day free trial, and nothing in it indicates independent audit or peer review. It does disclose audit counts, denominators and a negative result from the looser test, detail that allows parts of the analysis to be checked against itself.

A contested format under shifting enforcement

The research lands in an argument that has run for more than a year. PPC Land documented self-promotional listicles appearing as cited sources inside Google's AI Overviews in May 2025. After ranking fluctuations in January 2026, SISTRIX reported that sites using the tactic lost up to 49% of their visibility, with every identified case drawn from software-as-a-service and sharing traits such as scaled output, unedited AI text, artificial freshness updates and schema misuse. The most detailed single case, ClickUp's blog, lost 97.6% of its traffic from peak while carrying 249 competitor-alternatives listicles, all of them ranking ClickUp first.

Google extended its spam policies to AI Overviews and AI Mode on May 15, 2026, the first time its policy documentation explicitly addressed generated answers, including inauthentic mentions and scaled content abuse. It followed with a global spam update in June. Scrunch's post appeared on the day Google released its September 2026 spam update, the fourth of the year, with a rollout window of up to two weeks and no published target.

Whether the format has actually been demoted remains disputed. Kevin Indig's study of 5.32 million organic result rows, published in August, found that 62% of vendor-authored listicles had increased estimated organic traffic since January, with a median rise of 38%, and that vendor listicles appeared in the top 10 on 46.2% of business software results pages. The same memo cited Peec AI data showing no demotion of self-promoting listicles in citations across ChatGPT, Google's AI surfaces, Perplexity, Copilot and Gemini. Scrunch's finding that roughly three in 10 cited comparison pages are self-promoting sits comfortably alongside that.

There is also a mechanical reason the format keeps being retrieved. Peec AI's analysis of 5 million query fan-outs found ChatGPT inserting the words best and reviews, along with the current year, into its background searches even when users had typed none of them. That is the vocabulary of a "best [category]" page. Separate network-traffic analysis in August found that brands written into ChatGPT's own first search query reached the final answer 68.9% of the time, against 2.1% for brands that were merely retrieved. If that pattern generalises, much of the selection happens before any listicle is read.

Scrunch has published a run of citation studies this year. Its tracking of sponsored YouTube videos, covered by PPC Land earlier this week, recorded the citation rate falling from 18.4% to 10.5% within a single month between May and July 2026. Its analysis of source preferences over the same period found Reddit supplying 28.9% of ChatGPT's citations, while Google's AI surfaces cited YouTube at 21.1% and Facebook at 17%. Taken together, the studies describe citation patterns that differ by platform and move within weeks.

Why this matters for the marketing community

For the companies that have built content programmes around self-promoting lists, concentrated in business software, the numbers cut both ways. A cited list does move the author's odds, and it roughly doubles the chance of being named at all. Yet on every measure Scrunch publishes, the list's author is left out of the recommendation in the large majority of answers that cite it, and most of those answers recommend nobody.

Measurement is the second problem. Only 16% of brands systematically track AI visibility, according to the IAB framework, and Semrush found that only 36 of more than 1,200 tracked brands held top-100 mention status on every platform in every month of its study. A tactic whose effect depends heavily on the assistant, and whose apparent size depends on whether an analyst counts mentions or recommendations, is hard to evaluate through a single blended visibility score.

The evidence base for optimisation in general remains thin. A critical survey of 45 generative engine optimisation studies, posted to arXiv in July, found no technique producing a stable effect across platforms. Scrunch's looser test, which found nothing, is consistent with that caution. Its tighter test does detect an effect - a small one, on a sample the company itself describes as capped, from a vendor that sells the measurement.

The rival finding may prove the most consequential for smaller vendors, since the only scenario in which a competitor's lift survived testing involves a lower-visibility brand naming lower-visibility rivals. By Scrunch's own account, it is also the estimate least open to refinement. For now it stands as a single uncorrected result that points in a consistent direction without being precise.

Timeline

Summary

Who: Scrunch, an AI search visibility company that describes itself as a Sitecore company, published the study. The post carries the byline of Niharika Sharma, while Scrunch's press material credits Michael Ianelli, head of data science. The findings concern brands that publish self-promoting listicles, their named competitors, and seven AI answer surfaces: ChatGPT, Claude, Gemini, Microsoft Copilot, Perplexity, Google AI Overviews and Google AI Mode.

What: An analysis of roughly 10,000 AI-cited URLs found that 33% of classifiable cited pages were comparison or evaluation pages, of which roughly 30% were self-promoting listicles. When such a list was cited, the author brand's recommendation rate moved from about 4% to roughly 7% and its mention rate from about 19% to about 39%. A competitor was recommended instead of the author 24.3% of the time, with a causal rival effect surviving testing only when a lower-visibility brand named lower-visibility rivals.

When: Scrunch published the research on September 24, 2026. The underlying responses were collected between May 17 and June 30, 2026, with six of the seven surfaces effectively sampled from June 10 to 30.

Where: The study measured AI answers across seven assistant and search surfaces operated by OpenAI, Anthropic, Google, Microsoft and Perplexity, using Scrunch's monitored prompt set. Grok and Meta AI were excluded.

Why: Self-promoting listicles have become a common tactic for appearing in AI answers, while Google has extended its spam policies to generated surfaces and practitioners dispute whether the format has been demoted. The study offers a measured estimate of what a citation does for a list's author and its rivals, but the causal figures rest on about 6% of the eligible data from a vendor that sells AI visibility monitoring, and the mention effect has not yet passed the company's own falsification check.