Snoika
Reports · Research report

What Makes AI Engines Recommend a Brand: A Study of 16,000 Answers

Portrait of Anton VedeshinAnton VedeshinFounder, Snoika · 31 min read · updated 2026Statistics verified by Snoika

Abstract

Background: People now ask AI engines which product or service to choose, and the engines answer with brand names. It is not known whether the quality of pages behind those answers matter.

Objective: To test whether the quality of a webpage can impact decisions of ChatGPT, Perplexity, Gemini and Google AI Mode to recommend brands.

Methods. A pre-registered, observational study. We sent 4,000 US buyer questions from 40 industries once to each of the four engines. We found every recommended brand and the page linked behind each recommendation, then measured 53 technical and content properties of 21,097 pages and 9 properties of how each page presents each brand.

Results. The answers held 42,838 recommendations. Ten of the 53 page measures and all nine brand-on-page measures kept a link to recommendations, each in the same direction on all four engines. Pages with a year in the title backed extra recommendations at 1.66 times the rate of other pages. Newer, longer and faster pages also backed more. Within a page, a brand with its own H2 section had 6.7 times the odds of being the recommended one. The engines differ in where they link: ChatGPT linked to the brand's own site for 56% of its links, Perplexity for 8%.

Key findings

About half of the brands an answer names are recommended80,742 brand mentions, 42,838 recommendations (53%).
The engines link to very different placesChatGPT's links behind recommendations go to the brand's own site 56% of the time. Perplexity's go to publishers 79% of the time, and Gemini's 69%.
Listicles carry the recommendations54% of the measured pages are lists of options such as “best X”. Product pages follow at 13%.
Fresh pages back more recommendationsA year in the title goes with a rate 1.66 times as high, and older pages back fewer (rate ratio 0.84 per step in age).
Length, sections and speed matter a littleMore words, more H2 sections, a faster server and a faster first paint each keep a small link.
How a page presents a brand matters a lotThe brand named first on a page is the recommended one in 55% of cases; a brand named fourth or later, in 12%.
Many popular signals showed no linkA named author, a named reviewer, statistics per 100 words, first-person wording and most structured data types did not go with more recommendations.

Introduction

A buyer who once typed a search into Google and compared ten blue links now often asks an AI engine a full question: which CRM fits a ten-person agency, which mattress is best under $200. The answer is short, and it names brands. For a company, being one of the named brands is the new first page.

Much advice exists on how to earn that place, and little of it rests on open data. This study measures one part of the question in public: when an AI engine recommends a brand and links to a page, what is that page like, and does the page's quality go with how often it backs recommendations?

The thesis

The quality of a webpage impacts decisions of ChatGPT, Perplexity, Gemini and Google AI Mode to recommend brands.

“Quality” is not a score here. It is a set of 53 raw page measures, such as word count, heading structure, structured data, loading speed and page age, and 10 measures of how a page presents each brand it names. Each one is tested on its own.

The three research questions

RQ1Which technical and content measures of a page go with it backing more brand recommendations?
RQ2Does the way a page presents a brand (position, section, words spent, price, criticism) go with AI engines recommending that brand?
RQ3Where do the links behind AI recommendations point, and how often are recommended brands also cited?

RQ1 and RQ2 test the thesis. RQ3 describes the data.

Methods

This is an observational, cross-sectional study. Each prompt was sent once to each engine in one collection run, and nothing on any page was changed. The protocol, the prompt set, the settings and the code were committed to a public GitHub repository and tagged before collection started on 12 November 2026. The protocol follows the structure of the OSF preregistration template and the STROBE checklist for observational studies. This section describes the whole process from the first prompt to the final check.

Prompt set

The study uses 4,000 prompts: 40 industries, 5 subcategories in each, 5 question templates, 4 prompts per subcategory per template. The industries cover software, finance, consumer technology, apps and subscriptions, shopping, and travel and food. Local places such as restaurants were left out because their answers depend on the exact location, like the city or region.

Every prompt is a question whose only useful answer is one or more brands. The five templates are:

Best overall (“What is the best ...?”)
Best for a buyer
Must-have feature
Budget
A problem in the buyer's own words

No prompt names a brand, a year or a place. And the flow was as follows:

  1. 011We set up an automation in n8n to seed prompts.
  2. 022Claude Sonnet 5.5 wrote 8,000 candidate questions from a fixed seeding prompt.
  3. 033Code removed rule breaks and near-duplicates.
  4. 044Jev (version jev-1.14.0), an evaluation model that scores a yes-or-no question about a text from 0 to 1, then answered five questions about each candidate, and the four best per cell were kept.
  5. 055A person checked 200 random prompts by hand.

The first round of this check found that 14.4% of prompts could be answered with a type of product instead of a brand, so a brand rule was added and the whole set was seeded and filtered again. Both rounds are in the repository.

Engines and collection

All answers were collected from the United States, in English, logged out, on 12 and 13 November 2026. ChatGPT (with web search), Gemini and Google AI Mode were collected through the Oxylabs Web API.

Perplexity was collected through Perplexity's own Sonar API (model sonar, web search on), because collection of Perplexity through Oxylabs failed on 99 of 100 pilot prompts. Perplexity's answers therefore come from its API and not from its website, and are labelled so in the data.

Failed requests were retried up to two times. All 16,000 answers were collected. The prompts went out in four batches of ten industries because of storage limitations.

Finding recommendations

A brand list was built for each industry. Claude Sonnet 5.5 drafted names and aliases from the answers, products count for their maker, and fixed rules handle short aliases and everyday words. A person reviewed all 668 brands on the pilot lists. The full run added 6,609 more brand entries, drafted the same way and tidied by fixed rules; a person checked a random sample of 50 of them from the first batch and rejected none. Brand names were then matched in every answer. Being named is not being recommended, so each brand was labelled recommended or only mentioned with one rule, applied to every engine alike:

When the answer singles out its own picks (“the top choices are X and Y”, “I recommend X”), only those brands are recommended. Alternatives next to them count as mentioned.
When the answer singles out none and only gives a list, every brand it offers as an option is recommended.
Brands named for comparison, background or a warning are always only mentioned.

Jev answered three questions to apply this rule:

Whether the answer names its own pick
Whether each brand is one of its picks
Whether each brand is offered as an option

The link behind a recommendation is the URL cited in the same sentence, list item or paragraph as the brand. Each link was labelled by who owns the page relative to the recommended brand: own site, retailer, review platform, social media, other (government, education, Wikipedia), competitor, or publisher.

Pages and measures

Every website page behind a recommendation was fetched once. Pages that blocked the request or returned no text were fetched again through Oxylabs up to three tries. Pages were measured as they were fetched and never stored. Google's PageSpeed Insights (mobile) was run on every measured page.

Every measure is a raw value: a count, a length, seconds, or yes or no. There are no scores and no pass marks.

Technical measures (28)H1 count, skipped heading levels, title and meta description length, ten structured data measures and their count, speed from real Chrome users (LCP, INP, CLS, FCP, TTFB), lab speed (performance score and five timings), server response time, and whether the page offers a markdown version for LLMs.
Content measures (25)word count of the main content, H2 and H3 counts, lists, tables, internal and external links, broken links, statistics per 100 words, named author, named reviewer, page age, first-person wording, and the wording of the title and H1 (“best”, a year, a number).
Brand-on-page measures (10)for every brand a page names: its position among the page's brands, the H2 it first appears under, whether it has its own heading, the words in its section, its share of words, whether a price is shown, and whether the page criticizes it (judged by Jev).
Controlsdomain rank and page rank from a backlink index (0 to 1,000), page type (judged by Jev, 14 types), industry, and the brand's own domain rank as a measure of fame. Main content means the page without header, footer, navigation, sidebars and comments. Content measures use only pages with at least 200 words of main content.

Analysis

Each measure is tested on its own with the same controls, so the report can say which ones matter and by how much.

RQ1The outcome is a page's reach: the number of recommendations that link to it. A negative binomial regression models reach minus 1 on one measure plus domain rank, page rank, page type and industry, with standard errors clustered by domain.
RQ2A conditional logistic regression compares the brands named on one page behind one answer: which of them did the answer recommend? The model holds one measure plus brand fame, and uses pages that name at least two brands.
RQ3Shares with 95% bootstrap intervals (1,000 resamples, clustered by prompt).

A measure counts as a finding when its 95% interval excludes 1 after Benjamini-Hochberg correction at a false discovery rate of 5%, applied separately to RQ1 and RQ2. Both models were rerun for each engine.

As a robustness check, RQ1 was also fitted with all page measures in one model. Nothing was imputed: a page missing a measure leaves only that measure's model.

Quality checks

Automated labels can be wrong, so a person checked random samples against them: once in the pilot and once after the full run, on 30 answers and 30 pages drawn from all four batches. Each sample showed suggested marks from a second reader, which the person confirmed or changed. The protocol's rule is that a yes-or-no measure with agreement below kappa 0.7 is reported as exploratory only.

Table 1. Agreement between the automated values and a person's reading. Kappa is an agreement beyond chance. The protocol sets the bar at 0.7.

Recommended or only mentioned

Pilot
91.5% agreement, kappa 0.82 (130 brands, 30 answers)
Final check
92% agreement, kappa 0.82 (136 brands, 30 answers)

Link sits with the right brand

Pilot
not measured separately
Final check
150 of 150 links

Page type

Pilot
93%, kappa 0.92 (30 pages)
Final check
83%, kappa 0.75 (30 pages)

Page owner

Pilot
97%, kappa 0.96
Final check
78%, kappa 0.59 (41 brands)

Criticism

Pilot
98%, kappa 0.97 (61 brands, 20 pages, after the question was narrowed)
Final check
64%, kappa 0.36 (25 brands): reported as exploratory

Page measures (headings, words, links, dates, author, reviewer)

Pilot
Final check
330 of 330 values correct

Brand-on-page measures (position, section, words, price)

Pilot
Final check
108 of 108 values correct

Brand names matched in answers

Pilot
Final check
4 of 140 were everyday words matched as brands

Three results of the final check shape how this report should be read. The recommended-or-mentioned label is reliable overall, but less so on Perplexity (kappa 0.56), whose closing lines such as “choose X for this, Y for that” were labelled as only mentioned. The page owner label reached kappa 0.59: the domain rule missed some brands' own sites on other domains, so the own-site shares below are somewhat understated. And criticism fell to kappa 0.36, far below its pilot value, so it is reported as exploratory and kept out of the thesis test.

Changes after pre-registration

Two rules were added after the protocol was frozen, and both are recorded in it:

  • Fixed rules for tidying the brands added in the full run, with a sample check of the first batch, added when the first batch showed duplicate and everyday-word brands.
  • Social media pages were left out of the page measures, decided after the second batch, when 255 of 264 Reddit pages returned no text without a browser. They are still counted in the link statistics.

We moved criticism to exploratory as the rule from the protocol got applied.

Results

How to read the numbers

Rate ratio (RQ1)Every page in the sample backs at least one recommendation, so the model counts the extra ones. A rate ratio of 1.5 means pages with the property back extra recommendations at 1.5 times the rate of pages without it, with domain rank, page rank, page type and industry held equal. Below 1 means fewer.
Odds ratio (RQ2)The same idea for a yes-or-no outcome: the odds that a brand is the recommended one, compared with the other brands on the same page.
Per stepFor counts and continuous measures the ratio is per one standard deviation, a typical step between a middling page and a clearly higher one.
95% intervalThe range the true value plausibly lies in. An interval that excludes 1 after correction is a finding.
A link is not a causeThe study observes pages and does not change them in any way.

Data flow

All 4,000 prompts returned an answer on every engine. 15,954 of the 16,000 answers named at least one brand from the lists, and 15,723 held at least one recommendation.

4,000

What
prompts
Note
40 industries, 100 prompts each

16,000

What
answers
Note
each prompt sent once to four engines

80,742

What
brand mentions
Note
15,954 answers name at least one brand

42,838

What
recommendations
Note
53% of mentions; the rest are only mentioned

49,461

What
links behind recommendations
Note
23,137 different pages

22,085

What
website pages fetched
Note
left out: 941 social media pages, 111 video or Google result links

21,097

What
pages measured
Note
left out: 866 that returned no HTML, 122 not in English

18,940

What
pages in the content models
Note
left out: 2,157 with under 200 words of main content

Figure 1. From prompts to measured pages. Each row follows from the row above.

The engines are recommended at different rates. Per answer, Google AI Mode recommended 3.9 brands on average and Gemini 3.5, against 2.0 for ChatGPT and 1.4 for Perplexity. The reason is the style of the answer. ChatGPT and Perplexity name their own pick in about nine answers out of ten, while Google AI Mode does so in 18% and mostly gives a list (Figure 3).

Recommendations concentrate on a few pages. Of the 21,097 measured pages, 13,174 (62%) back exactly one recommendation, 7,923 back more than one, and 552 back ten or more. The measured pages sit on 7,369 different sites.

Where the links point (RQ3)

The engines differ most in whose page they link to. ChatGPT sends 56% of its links behind recommendations to the recommended brand's own site. Google AI Mode does so for 26%, Gemini for 9% and Perplexity for 8%. Perplexity and Gemini rely on publishers instead: media, magazines and blogs take 79% and 69% of their links. A competitor's page backs between 4% and 14% of recommendations, depending on the engine.

Stacked bars by engine: ChatGPT links to own sites 56%, Perplexity to publishers 79%
Figure 2. Who owns the page behind a recommendation, by engine. The three smaller types are retailer, review platform and other (government, education, Wikipedia); Table A1 gives every share.

A recommended brand's own site is cited somewhere in the answer for 64% of ChatGPT's recommendations, but for only 16% to 29% on the other three engines. Many recommendations carry no link at all: 40% on Google AI Mode, 37% on Gemini and 19% on ChatGPT. Perplexity links almost every recommendation.

Bars by engine: own site cited, no link, and answers naming their own pick
Figure 3. Three shares by engine: how often a recommended brand has its own site cited, how often it has no link, and how often the answer names its own pick.

The industry matters as well. In software industries a recommended brand's own site is cited for about 40% to 45% of recommendations; in consumer goods such as beauty, home appliances and headphones it is 10% to 12% (Figure 4). Engines draw more on vendor sites for software questions and more on third-party pages for product questions.

Bars by industry: own site cited for 45% in software, 10% in beauty and skincare
Figure 4. Share of recommended brands whose own site is cited anywhere in the answer, by industry, all engines together.

More than half of the measured pages (54%) are listicles: pages that present several options, such as “best X” or “X alternatives”. Product pages (13%), articles (9%) and homepages (8%) follow.

Bar chart of page types: listicles 54%, product pages 13%, articles 8.7%
Figure 5. Page type of the 21,097 measured pages behind recommendations.

Social media is a small part of the links. 941 of the 23,137 pages behind recommendations were on social platforms, 713 of them on Reddit, and almost all came from Gemini and Google AI Mode.

Table 2. The ten sites whose pages back the most recommendations.

nerdwallet.com

Pages measured
275
Recommendations backed
1,098

pcmag.com

Pages measured
230
Recommendations backed
1,082

cnet.com

Pages measured
153
Recommendations backed
887

forbes.com

Pages measured
258
Recommendations backed
869

cnbc.com

Pages measured
175
Recommendations backed
755

rtings.com

Pages measured
176
Recommendations backed
710

nytimes.com

Pages measured
132
Recommendations backed
604

zapier.com

Pages measured
63
Recommendations backed
436

rankite.com

Pages measured
228
Recommendations backed
420

usnews.com

Pages measured
115
Recommendations backed
281

Page measures (RQ1)

Ten of the 53 page measures kept a link to reach after correction. Seven go with more recommendations and three with fewer (Figure 6, Table 3).

Rate ratios for 53 page measures; year and best in title or H1 show the largest links
Figure 6. All 53 page measures: rate ratio with 95% interval, on a log scale. Purple marks kept a link after correction.

Table 3. The ten page measures that kept a link to reach after correction. For yes-or-no measures the ratio compares yes with no; for the others it is per one standard deviation.

Title contains a year

Rate ratio
1.66
95% interval
1.52 to 1.81
Interpretation
Pages with a year in the title back extra recommendations at 1.66 times the rate of pages without one

H1 contains a year

Rate ratio
1.50
95% interval
1.34 to 1.67
Interpretation
The same pattern for a year in the main heading

Title contains “best”

Rate ratio
1.46
95% interval
1.33 to 1.61
Interpretation
Pages titled as a “best of” back more recommendations

H1 contains “best”

Rate ratio
1.35
95% interval
1.22 to 1.50
Interpretation
The same pattern for the main heading

Schema: FAQPage

Rate ratio
1.15
95% interval
1.04 to 1.26
Interpretation
Pages with FAQ structured data back slightly more

H2 sections (count)

Rate ratio
1.10
95% interval
1.05 to 1.15
Interpretation
More sections, more recommendations (7 to 17 H2 headings)

Word count

Rate ratio
1.10
95% interval
1.04 to 1.15
Interpretation
Longer pages back more (1,840 to 4,470 words)

TTFB, server wait (real users)

Rate ratio
0.94
95% interval
0.90 to 0.98
Interpretation
A slower server goes with fewer recommendations (0.54 seconds)

FCP, first paint (real users)

Rate ratio
0.93
95% interval
0.89 to 0.97
Interpretation
A slower first paint goes with fewer (0.82 seconds)

Page age (days since update)

Rate ratio
0.84
95% interval
0.80 to 0.88
Interpretation
Older pages back fewer (about 54 to 287 days since the last update)

The strongest links are about what the page says it is. A year in the title or in the H1, and the word “best” in either, go with 35% to 66% higher rates. These hold with page type held equal, so they are not only an effect of listicles being cited more. Among pages with at least 200 words, 46% of those with a year in the title back more than one recommendation, against 31% of those without.

Freshness shows a second time in page age. Each step toward an older page lowers the rate by 16%. Length and structure keep smaller links, about 10% per step for word count and for H2 sections. Two speed measures from real users, time to first byte and first contentful paint, keep links of 6% to 7% per step. FAQPage structured data keeps a link of 15%.

The other 43 measures showed no link after correction. Among them are several that are often advised: a named author, a named reviewer, statistics per 100 words, first-person wording, tables and lists, external links, the number of H1 headings, and nine of the ten structured data measures. The lab speed scores showed none either. Two measures came close and are not findings: largest contentful paint from real users (rate ratio 0.95) and a markdown version of the page (1.16).

Brand on a page (RQ2)

RQ2 looks inside the page. The comparison is between brands that have the same page names for the same answer: 26,490 such groups with 234,208 brand rows, a median of 8 brands per group. On average 21% of the brands in a group were recommended by the answer.

All nine tested measures kept a link, and the links are large.

Odds ratios for brand-on-page measures; own H2 heading highest, later position lowest
Figure 7. Brand-on-page measures: odds ratio with 95% interval, on a log scale. Criticism is shown for completeness and is exploratory.

Table 4. The nine tested brand-on-page measures. Each compares brands named on the same page behind the same answer.

Brand has its own H2 heading

Odds ratio
6.7
95% interval
6.41 to 7.03
In plain words
A brand with its own H2 section has 6.7 times the odds of being the recommended one

Brand has its own H3 heading

Odds ratio
4.9
95% interval
4.75 to 5.12
In plain words
The same pattern for an H3 section

Price shown in the brand's section

Odds ratio
4.4
95% interval
4.25 to 4.54
In plain words
Brands whose section shows a price or plan are recommended more

Words relative to other brands

Odds ratio
3.2
95% interval
2.80 to 3.68
In plain words
More words than the other brands on the page; unstable across engines

Words in the brand's section

Odds ratio
3.1
95% interval
3.03 to 3.15
In plain words
More words in the brand's section, more often recommended

Share of the page's brand words

Odds ratio
1.9
95% interval
1.86 to 1.91
In plain words
A larger share of the page's brand words

Position as a share of all brands

Odds ratio
0.42
95% interval
0.41 to 0.42
In plain words
Brands named later in the page are recommended less

Position among the page's brands

Odds ratio
0.37
95% interval
0.36 to 0.37
In plain words
The same pattern for the raw position

Which H2 the brand first appears under

Odds ratio
0.32
95% interval
0.31 to 0.32
In plain words
Brands that first appear under a later H2 are recommended less

The brand a page names first is the recommended one in 55% of cases, and a brand named fourth or later in 12%. Brands with their own H2 section are recommended in 33% of cases against 18% for the others, and brands with a price in their section in 29% against 17%.

One measure is unstable. The ratio of a brand's words to the other brands' words has a pooled odds ratio of 3.2, but it ranges from 1.02 on Perplexity to 90 on ChatGPT, with a very wide interval there. Its direction is the same everywhere, so its size should not be relied on.

Results by engine

The thesis rule asks for the same direction on at least three engines. All ten RQ1 findings and all nine RQ2 findings point the same way on all four.

The size differs by engine. For the ten RQ1 findings, the interval excludes 1 for all ten on Google AI Mode, for 8 on Gemini, for 6 on Perplexity and for none on ChatGPT. On ChatGPT every effect has the same direction but is smaller, and its intervals include 1. The page measures therefore tell most about Google's two engines and Perplexity, and least about ChatGPT, which mostly links to brands' own sites.

Rate ratios for the ten RQ1 findings on ChatGPT, Gemini, Google AI Mode and Perplexity
Figure 8. The ten RQ1 findings on each engine. Reach is counted on that engine only.

For RQ2, every tested measure has an interval that excludes 1 on every engine (Table A5).

Robustness

RQ1 was fitted again with all page measures in one model, so each measure competes with the others. Measures available for under 90% of pages (the real-user speed measures and page age) cannot enter this model, and of any two measures correlated above 0.8 one is kept, which removes “best” in the H1 in favour of “best” in the title. Of the six findings that entered, five held: word count, H2 sections (count), title contains “best”, title contains a year, H1 contains a year. FAQPage structured data did not: its rate ratio fell to 1.07 with an interval that includes 1, so its link is the weakest of the ten.

Exploratory: criticism

Brands that a page criticizes were recommended more often than the other brands on the same page: odds ratio 4.0 (95% interval 3.89 to 4.14), in the same direction on every engine. This result is exploratory. The criticism label agreed with a person's reading at only kappa 0.36 in the final check, so the label itself is unreliable, and the result is outside the correction and the thesis test. A likely reading is that pages write the most, good and bad, about the brands they cover in depth, and those are the brands the engines recommend. The study cannot confirm it.

Discussion

Under the rule fixed before collection, the thesis is supported: 19 measures kept a link to recommendations after correction, each in the same direction on all four engines. What “quality” turned out to mean is narrower than the word suggests. The page properties that go with recommendations are freshness, a title that matches the buyer's question, length, clear sections and speed. Signals of authorship and most markup showed nothing.

What this means for a website

The points below are readings of the links above.

Keep pages current, and show itPage age and a year in the title are among the strongest links, and 29% of the measured pages show no date at all.
Give each product or brand its own sectionA heading of its own, enough words and a visible price all go with being the recommended brand on a page.
Be early on the pages that list youPosition on the page has one of the largest links in the study. For a brand, the order in third-party lists is not a detail.
Know which engine you are aiming atChatGPT recommends from brands' own sites; Perplexity and Gemini recommend from publishers. Visibility on Perplexity and Gemini is likely to depend on appearing in publishers' lists, and visibility on ChatGPT more on your own pages.
Know your industrySoftware brands have their own sites cited about four times as often as beauty or home appliance brands.
Treat speed as hygieneServer wait and first paint keep small links. Lab scores do not.
Do not expect markup or bylines to carry a pageAuthor and reviewer names, statistic density and most structured data types showed no link here.

What the study does not show

CauseThe study did not change any page. A year in the title may raise a page's chance of being used, or pages that engines already favour may simply be the kind that put years in titles.
Search matchingMost prompts ask for “the best” option. Engines search the web before answering, and a page titled “best X” matches that search. The title findings may describe retrieval as much as a judgment of quality.
Pages that are never citedEvery page in the sample backs at least one recommendation. The study compares pages cited once with pages cited often. It cannot say what separates a cited page from an uncited one.
How much of RQ2 is built inAn engine often cites a page for the passage about one brand, and the study attaches a link to the brand it sits next to. The brand with its own section is then likely to be the recommended one. The RQ2 links are real patterns, but part of their size comes from how citation works.

Limitations

One market and one runAnswers change over time and by place. This is one collection, from the United States, in English, over two days.
Logged-out answersLogged-in and personalised answers can differ. Perplexity was collected through its API.
Automated judgmentsBrand matching, the recommended-or-mentioned label, page type and page owner are automated. Their measured error is in Table 1: about 3% of matched brand names were everyday words, the page owner label reached kappa 0.59, and the recommended label is weaker on Perplexity. Jev can also give a different label when asked again, two reruns on pilot answers changed about 2% of labels.
Link attributionA citation in a part of the answer that names no brand is given to the brand whose list item or section it sits in, which can be wrong. In the final check all 150 links sat with the right brand.
Short pagesIn the final check the study had recorded no brand-on-page row for 14 of the 41 recommended brands on the sampled pages, 10 of them on pages with under 400 words of main content. RQ2 rests on the pages where brands could be located.
Speed dataReal-user speed exists only for pages and sites with enough Chrome traffic; the rest use site-level or lab data.
TimingPages were measured when fetched, shortly after the answers were collected, not at the moment the engine read them.
Social media941 social media pages were left out of the page measures.

Conclusion

Across 16,000 answers from four AI engines, the pages behind brand recommendations are not interchangeable. Pages that are fresh, say what they are in the title, cover their subject at length in clear sections and load fast back more recommendations, and within a page the brand that comes first and has its own section is the one most often recommended. The links hold in the same direction on ChatGPT, Gemini, Google AI Mode and Perplexity, though they are weakest on ChatGPT. The thesis that page quality goes with AI recommendations is supported under the pre-registered rule.

Data and code

The protocol, the 4,000 prompts, the settings, the collection workflows, the analysis script, the dataset and the result tables are public:

github.com/andrii-dolinskyi/ai-recommendation-study

Pre-registrationtag preregistration, committed on 12 November 2026 before collection.
Releasedprompts, recommendations and mentions with their links, answer flags, page measures, brand-on-page measures, PageSpeed results, ranks, the quality check files and every result table in this report.
Not releasedAPI keys, the engines' answer texts and copies of other websites' pages.
ToolsOxylabs Web API, Perplexity Sonar API, Jev jev-1.14.0, Claude Sonnet 5.5 (claude-sonnet-5-5), Google PageSpeed Insights, n8n, and Python with statsmodels (random seed 20261105).

How to cite. Vedeshin, A. (2026). What Makes AI Engines Recommend a Brand: A Study of 16,000 Answers. Snoika Reports.

Appendix

Table A1. Share of links behind recommendations by page owner, with 95% bootstrap intervals (1,000 resamples, clustered by prompt).

Own site

ChatGPT
55.9% (54.6 to 57.1)
Gemini
9.4% (8.8 to 10.0)
Google AI Mode
25.8% (24.7 to 26.7)
Perplexity
8.2% (7.6 to 8.8)

Publisher

ChatGPT
33.8% (32.5 to 35.1)
Gemini
68.9% (67.7 to 70.2)
Google AI Mode
50.6% (49.2 to 52.0)
Perplexity
78.5% (77.4 to 79.6)

Competitor

ChatGPT
3.6% (3.2 to 4.1)
Gemini
14.4% (13.5 to 15.4)
Google AI Mode
12.1% (11.2 to 13.1)
Perplexity
10.8% (10.1 to 11.6)

Social media

ChatGPT
0.0% (0.0 to 0.1)
Gemini
5.7% (5.1 to 6.3)
Google AI Mode
8.0% (7.3 to 8.7)
Perplexity
0.1% (0.0 to 0.1)

Retailer

ChatGPT
5.1% (4.5 to 5.7)
Gemini
0.9% (0.7 to 1.1)
Google AI Mode
2.4% (2.0 to 2.8)
Perplexity
1.3% (1.1 to 1.6)

Review platform

ChatGPT
0.7% (0.5 to 0.9)
Gemini
0.6% (0.4 to 0.8)
Google AI Mode
0.8% (0.6 to 1.0)
Perplexity
0.9% (0.7 to 1.1)

Other

ChatGPT
0.8% (0.6 to 1.1)
Gemini
0.0% (0.0 to 0.1)
Google AI Mode
0.3% (0.2 to 0.4)
Perplexity
0.1% (0.1 to 0.2)

Table A2. Three shares by engine, with 95% bootstrap intervals.

Recommended brands whose own site is cited anywhere in the answer

ChatGPT
63.8% (62.3 to 65.2)
Gemini
16.3% (15.6 to 17.1)
Google AI Mode
28.7% (27.6 to 29.9)
Perplexity
20.3% (19.1 to 21.5)

Recommended brands with no link

ChatGPT
19.5% (18.3 to 20.8)
Gemini
36.7% (35.5 to 37.9)
Google AI Mode
39.9% (38.7 to 41.0)
Perplexity
0.8% (0.4 to 1.1)

Answers that name their own pick

ChatGPT
91.0% (90.0 to 91.8)
Gemini
42.6% (41.0 to 44.1)
Google AI Mode
18.4% (17.3 to 19.6)
Perplexity
91.2% (90.3 to 92.1)

Table A3. All 53 page measures (RQ1), pooled over the four engines. q is the p-value after Benjamini-Hochberg correction.

H1 headings (count)

Family
Technical
Pages
21,095
Rate ratio
0.99
95% interval
0.95 to 1.03
p
0.635
q
0.729
Finding

Skipped heading levels

Family
Technical
Pages
21,095
Rate ratio
0.97
95% interval
0.93 to 1.01
p
0.201
q
0.453
Finding

Title length

Family
Technical
Pages
21,095
Rate ratio
0.97
95% interval
0.93 to 1.01
p
0.149
q
0.390
Finding

Meta description length

Family
Technical
Pages
21,095
Rate ratio
0.98
95% interval
0.95 to 1.02
p
0.387
q
0.590
Finding

Schema types (count)

Family
Technical
Pages
21,095
Rate ratio
1.03
95% interval
0.99 to 1.08
p
0.139
q
0.390
Finding

Schema: Article

Family
Technical
Pages
21,095
Rate ratio
1.07
95% interval
0.93 to 1.23
p
0.368
q
0.590
Finding

Schema: Product

Family
Technical
Pages
21,095
Rate ratio
1.11
95% interval
0.93 to 1.31
p
0.242
q
0.494
Finding

Schema: Review

Family
Technical
Pages
21,095
Rate ratio
1.10
95% interval
0.91 to 1.32
p
0.316
q
0.590
Finding

Schema: FAQPage

Family
Technical
Pages
21,095
Rate ratio
1.15
95% interval
1.04 to 1.26
p
0.005
q
0.026
Finding
Yes

Schema: ItemList

Family
Technical
Pages
21,095
Rate ratio
0.96
95% interval
0.83 to 1.12
p
0.608
q
0.729
Finding

Schema: HowTo

Family
Technical
Pages
21,095
Rate ratio
1.12
95% interval
0.80 to 1.57
p
0.504
q
0.703
Finding

Schema: Organization

Family
Technical
Pages
21,095
Rate ratio
1.05
95% interval
0.96 to 1.15
p
0.329
q
0.590
Finding

Schema: Person

Family
Technical
Pages
21,095
Rate ratio
1.06
95% interval
0.97 to 1.15
p
0.220
q
0.466
Finding

Schema: BreadcrumbList

Family
Technical
Pages
21,095
Rate ratio
0.99
95% interval
0.91 to 1.08
p
0.820
q
0.870
Finding

Schema: other types

Family
Technical
Pages
21,095
Rate ratio
0.98
95% interval
0.89 to 1.07
p
0.622
q
0.729
Finding

Server response time (lab)

Family
Technical
Pages
20,908
Rate ratio
0.99
95% interval
0.96 to 1.02
p
0.594
q
0.729
Finding

LCP (real users)

Family
Technical
Pages
18,916
Rate ratio
0.95
95% interval
0.92 to 0.99
p
0.012
q
0.055
Finding

INP (real users)

Family
Technical
Pages
17,811
Rate ratio
1.00
95% interval
0.98 to 1.03
p
0.770
q
0.833
Finding

CLS (real users)

Family
Technical
Pages
19,009
Rate ratio
1.02
95% interval
0.98 to 1.05
p
0.372
q
0.590
Finding

FCP, first paint (real users)

Family
Technical
Pages
19,021
Rate ratio
0.93
95% interval
0.89 to 0.97
p
<0.001
q
0.005
Finding
Yes

TTFB, server wait (real users)

Family
Technical
Pages
18,650
Rate ratio
0.94
95% interval
0.90 to 0.98
p
0.004
q
0.025
Finding
Yes

Performance score (lab)

Family
Technical
Pages
20,499
Rate ratio
0.96
95% interval
0.92 to 1.00
p
0.034
q
0.140
Finding

LCP (lab)

Family
Technical
Pages
20,503
Rate ratio
1.01
95% interval
0.97 to 1.06
p
0.541
q
0.729
Finding

CLS (lab)

Family
Technical
Pages
21,017
Rate ratio
1.02
95% interval
0.99 to 1.06
p
0.154
q
0.390
Finding

Total Blocking Time (lab)

Family
Technical
Pages
20,503
Rate ratio
0.99
95% interval
0.96 to 1.02
p
0.572
q
0.729
Finding

FCP (lab)

Family
Technical
Pages
20,623
Rate ratio
1.03
95% interval
1.00 to 1.07
p
0.089
q
0.317
Finding

Speed Index (lab)

Family
Technical
Pages
20,619
Rate ratio
0.96
95% interval
0.92 to 1.01
p
0.096
q
0.317
Finding

Markdown version of the page

Family
Technical
Pages
21,095
Rate ratio
1.16
95% interval
1.02 to 1.33
p
0.028
q
0.122
Finding

Word count

Family
Content
Pages
18,939
Rate ratio
1.10
95% interval
1.04 to 1.15
p
<0.001
q
0.002
Finding
Yes

H2 sections (count)

Family
Content
Pages
18,939
Rate ratio
1.10
95% interval
1.05 to 1.15
p
<0.001
q
<0.001
Finding
Yes

H3 sections (count)

Family
Content
Pages
18,939
Rate ratio
1.04
95% interval
0.99 to 1.08
p
0.094
q
0.317
Finding

Lists (count)

Family
Content
Pages
18,939
Rate ratio
0.98
95% interval
0.94 to 1.02
p
0.381
q
0.590
Finding

List items (count)

Family
Content
Pages
18,939
Rate ratio
0.99
95% interval
0.95 to 1.03
p
0.646
q
0.729
Finding

Tables (count)

Family
Content
Pages
18,939
Rate ratio
1.00
95% interval
0.95 to 1.05
p
0.991
q
0.991
Finding

External links

Family
Content
Pages
18,939
Rate ratio
0.96
95% interval
0.91 to 1.02
p
0.168
q
0.405
Finding

Internal links

Family
Content
Pages
18,939
Rate ratio
1.04
95% interval
0.99 to 1.09
p
0.110
q
0.342
Finding

Broken links (count)

Family
Content
Pages
18,939
Rate ratio
0.98
95% interval
0.95 to 1.02
p
0.365
q
0.590
Finding

Broken links (share)

Family
Content
Pages
17,254
Rate ratio
1.03
95% interval
0.99 to 1.07
p
0.205
q
0.453
Finding

Statistics per 100 words

Family
Content
Pages
18,939
Rate ratio
0.98
95% interval
0.95 to 1.02
p
0.389
q
0.590
Finding

Author named

Family
Content
Pages
18,939
Rate ratio
1.04
95% interval
0.94 to 1.16
p
0.445
q
0.648
Finding

Reviewer named

Family
Content
Pages
18,939
Rate ratio
1.01
95% interval
0.86 to 1.20
p
0.892
q
0.916
Finding

Page age (days since update)

Family
Content
Pages
13,922
Rate ratio
0.84
95% interval
0.80 to 0.88
p
<0.001
q
<0.001
Finding
Yes

No date shown

Family
Content
Pages
18,939
Rate ratio
0.94
95% interval
0.84 to 1.05
p
0.288
q
0.565
Finding

First person singular per 100 words

Family
Content
Pages
18,939
Rate ratio
1.01
95% interval
0.98 to 1.05
p
0.453
q
0.648
Finding

First person plural per 100 words

Family
Content
Pages
18,939
Rate ratio
1.00
95% interval
0.96 to 1.03
p
0.898
q
0.916
Finding

Title contains “best”

Family
Content
Pages
18,939
Rate ratio
1.46
95% interval
1.33 to 1.61
p
<0.001
q
<0.001
Finding
Yes

Title contains a year

Family
Content
Pages
18,939
Rate ratio
1.66
95% interval
1.52 to 1.81
p
<0.001
q
<0.001
Finding
Yes

Title contains a number

Family
Content
Pages
18,939
Rate ratio
0.98
95% interval
0.89 to 1.07
p
0.646
q
0.729
Finding

H1 contains “best”

Family
Content
Pages
18,939
Rate ratio
1.35
95% interval
1.22 to 1.50
p
<0.001
q
<0.001
Finding
Yes

H1 contains a year

Family
Content
Pages
18,939
Rate ratio
1.50
95% interval
1.34 to 1.67
p
<0.001
q
<0.001
Finding
Yes

H1 contains a number

Family
Content
Pages
18,939
Rate ratio
0.98
95% interval
0.90 to 1.07
p
0.618
q
0.729
Finding

H1 contains “I”

Family
Content
Pages
18,939
Rate ratio
0.95
95% interval
0.73 to 1.23
p
0.687
q
0.759
Finding

H1 contains “we”

Family
Content
Pages
18,939
Rate ratio
1.13
95% interval
0.96 to 1.34
p
0.154
q
0.390
Finding

Table A4. The ten RQ1 findings on each engine: rate ratio with 95% interval. Grey values have an interval that includes 1.

Title contains a year

ChatGPT
1.21 (0.92 to 1.59)
Gemini
1.52 (1.36 to 1.70)
Google AI Mode
1.66 (1.41 to 1.95)
Perplexity
1.59 (1.40 to 1.81)

H1 contains a year

ChatGPT
1.23 (0.93 to 1.62)
Gemini
1.45 (1.29 to 1.61)
Google AI Mode
1.55 (1.34 to 1.79)
Perplexity
1.25 (1.00 to 1.56)

Title contains “best”

ChatGPT
1.30 (1.00 to 1.69)
Gemini
1.43 (1.26 to 1.61)
Google AI Mode
1.58 (1.36 to 1.83)
Perplexity
1.46 (1.25 to 1.69)

H1 contains “best”

ChatGPT
1.30 (1.00 to 1.69)
Gemini
1.23 (1.09 to 1.39)
Google AI Mode
1.51 (1.30 to 1.75)
Perplexity
1.38 (1.18 to 1.62)

Schema: FAQPage

ChatGPT
1.13 (0.95 to 1.35)
Gemini
1.22 (1.10 to 1.34)
Google AI Mode
1.17 (1.05 to 1.31)
Perplexity
1.17 (1.00 to 1.38)

H2 sections (count)

ChatGPT
1.02 (0.94 to 1.12)
Gemini
1.15 (1.08 to 1.22)
Google AI Mode
1.13 (1.08 to 1.19)
Perplexity
1.19 (1.08 to 1.30)

Word count

ChatGPT
1.08 (0.99 to 1.18)
Gemini
1.15 (1.06 to 1.24)
Google AI Mode
1.13 (1.07 to 1.19)
Perplexity
1.17 (1.05 to 1.30)

TTFB, server wait (real users)

ChatGPT
0.99 (0.91 to 1.08)
Gemini
0.97 (0.92 to 1.02)
Google AI Mode
0.92 (0.87 to 0.97)
Perplexity
0.97 (0.90 to 1.06)

FCP, first paint (real users)

ChatGPT
0.97 (0.90 to 1.06)
Gemini
0.96 (0.91 to 1.01)
Google AI Mode
0.93 (0.87 to 0.99)
Perplexity
0.95 (0.88 to 1.03)

Page age (days since update)

ChatGPT
0.90 (0.81 to 1.01)
Gemini
0.85 (0.80 to 0.91)
Google AI Mode
0.84 (0.80 to 0.89)
Perplexity
0.82 (0.77 to 0.86)

Table A5. Brand-on-page measures (RQ2) on each engine: odds ratio with 95% interval.

Position among the page's brands

ChatGPT
0.28 (0.27 to 0.30)
Gemini
0.41 (0.40 to 0.42)
Google AI Mode
0.37 (0.36 to 0.38)
Perplexity
0.37 (0.37 to 0.38)

Position as a share of all brands

ChatGPT
0.36 (0.35 to 0.38)
Gemini
0.45 (0.44 to 0.46)
Google AI Mode
0.44 (0.43 to 0.46)
Perplexity
0.40 (0.39 to 0.41)

Which H2 the brand first appears under

ChatGPT
0.26 (0.25 to 0.27)
Gemini
0.36 (0.35 to 0.38)
Google AI Mode
0.33 (0.31 to 0.34)
Perplexity
0.30 (0.29 to 0.31)

Brand has its own H2 heading

ChatGPT
10.07 (9.04 to 11.22)
Gemini
5.55 (5.08 to 6.06)
Google AI Mode
6.60 (6.02 to 7.25)
Perplexity
6.06 (5.56 to 6.60)

Brand has its own H3 heading

ChatGPT
3.51 (3.19 to 3.87)
Gemini
5.52 (5.15 to 5.91)
Google AI Mode
5.16 (4.77 to 5.58)
Perplexity
5.00 (4.66 to 5.37)

Words in the brand's section

ChatGPT
3.41 (3.24 to 3.59)
Gemini
2.96 (2.86 to 3.06)
Google AI Mode
3.06 (2.95 to 3.18)
Perplexity
2.99 (2.88 to 3.10)

Share of the page's brand words

ChatGPT
2.09 (2.02 to 2.17)
Gemini
2.04 (1.98 to 2.10)
Google AI Mode
2.23 (2.14 to 2.33)
Perplexity
1.62 (1.59 to 1.66)

Words relative to other brands

ChatGPT
90.10 (39.00 to 208.16)
Gemini
2.83 (2.66 to 3.01)
Google AI Mode
4.32 (3.87 to 4.81)
Perplexity
1.02 (1.00 to 1.03)

Price shown in the brand's section

ChatGPT
5.34 (4.84 to 5.88)
Gemini
4.53 (4.27 to 4.80)
Google AI Mode
4.72 (4.40 to 5.07)
Perplexity
3.80 (3.59 to 4.02)

Page criticizes the brand (exploratory)

ChatGPT
3.58 (3.29 to 3.90)
Gemini
4.04 (3.83 to 4.27)
Google AI Mode
4.53 (4.24 to 4.85)
Perplexity
3.86 (3.65 to 4.08)

Table A6. Robustness: the RQ1 findings that could enter the single model with all page measures (16,699 pages, 45 measures).

Schema: FAQPage

Rate ratio, all measures in one model
1.07
95% interval
0.97 to 1.18

Word count

Rate ratio, all measures in one model
1.10
95% interval
1.03 to 1.17

H2 sections (count)

Rate ratio, all measures in one model
1.07
95% interval
1.03 to 1.12

Title contains “best”

Rate ratio, all measures in one model
1.24
95% interval
1.13 to 1.37

Title contains a year

Rate ratio, all measures in one model
1.39
95% interval
1.24 to 1.57

H1 contains a year

Rate ratio, all measures in one model
1.20
95% interval
1.07 to 1.34

Table A7. Median and quartiles of every numeric page measure. Times are in seconds unless the name says otherwise.

H1 headings (count)

Pages
21,097
25%
1.00
Median
1.00
75%
1.00

Skipped heading levels

Pages
21,097
25%
0.00
Median
0.00
75%
1.00

Title length

Pages
21,097
25%
48.0
Median
57.0
75%
67.0

Meta description length

Pages
21,097
25%
131
Median
148
75%
159

Schema types (count)

Pages
21,097
25%
4.00
Median
9.00
75%
12.0

Server response time (lab)

Pages
20,910
25%
6.00
Median
15.0
75%
68.0

LCP (real users)

Pages
18,918
25%
1.58
Median
2.05
75%
2.69

INP (real users)

Pages
17,813
25%
127
Median
171
75%
228

CLS (real users)

Pages
19,011
25%
0.00
Median
0.02
75%
0.06

FCP, first paint (real users)

Pages
19,023
25%
1.32
Median
1.73
75%
2.26

TTFB, server wait (real users)

Pages
18,652
25%
0.63
Median
0.86
75%
1.23

Performance score (lab)

Pages
20,501
25%
42.0
Median
61.0
75%
82.0

LCP (lab)

Pages
20,505
25%
2.38
Median
5.38
75%
10.9

CLS (lab)

Pages
21,019
25%
0.00
Median
0.00
75%
0.04

Total Blocking Time (lab)

Pages
20,505
25%
136
Median
536
75%
1,383

FCP (lab)

Pages
20,625
25%
1.35
Median
1.99
75%
2.90

Speed Index (lab)

Pages
20,621
25%
3.63
Median
5.77
75%
8.71

Word count

Pages
21,097
25%
718
Median
1,785
75%
3,278

H2 sections (count)

Pages
21,097
25%
3.00
Median
7.00
75%
12.0

H3 sections (count)

Pages
21,097
25%
1.00
Median
8.00
75%
16.0

Lists (count)

Pages
21,097
25%
1.00
Median
4.00
75%
13.0

List items (count)

Pages
21,097
25%
3.00
Median
17.0
75%
50.0

Tables (count)

Pages
21,097
25%
0.00
Median
0.00
75%
1.00

External links

Pages
21,097
25%
0.00
Median
2.00
75%
12.0

Internal links

Pages
21,097
25%
5.00
Median
15.0
75%
32.0

Broken links (count)

Pages
21,097
25%
0.00
Median
0.00
75%
0.00

Broken links (share)

Pages
18,184
25%
0.00
Median
0.00
75%
0.00

Statistics per 100 words

Pages
20,509
25%
0.95
Median
2.11
75%
3.93

Page age (days since update)

Pages
14,899
25%
16.0
Median
59.0
75%
182

First person singular per 100 words

Pages
20,509
25%
0.00
Median
0.05
75%
0.29

First person plural per 100 words

Pages
20,509
25%
0.09
Median
0.38
75%
0.90

Table A8. Share of measured pages with each of the main yes-or-no properties.

Author named

Share of measured pages
78%

Title contains “best”

Share of measured pages
49%

Title contains a year

Share of measured pages
45%

H1 contains “best”

Share of measured pages
46%

H1 contains a year

Share of measured pages
37%

Schema: FAQPage

Share of measured pages
35%

No date shown

Share of measured pages
29%

Markdown version of the page

Share of measured pages
9%

Reviewer named

Share of measured pages
8%

Table A9. Social media pages behind recommendations, counted and left out of the page measures.

reddit.com

Pages behind recommendations
713

youtube.com

Pages behind recommendations
114

facebook.com

Pages behind recommendations
54

medium.com

Pages behind recommendations
25

linkedin.com

Pages behind recommendations
17

instagram.com

Pages behind recommendations
11

tiktok.com

Pages behind recommendations
6

pinterest.com

Pages behind recommendations
1

See how AI talks about your brand

Get free audit