Commercial content can surface in AI answers. That is not yet a recommendation effect.

Commercial and self-promotional pages are showing up in AI answers.

That finding matters. But it is easy to make it mean too much.

A page can appear as a source without its brand being named. A brand can be named without being recommended. A recommendation can be ignored. A consumer can accept help with research while refusing to let an AI make the final choice.

Several recent publisher studies now give us evidence at different points along that path. Peec AI examined 232,000 citations. DataPulse and GetCited scanned more than 100,000 matched English citations for disclosed commercial markers. Scrunch compared answers that did and did not cite an existing self-promotional listicle. Ahrefs tracked 34 pages promoting two of its own brands.[1][2][3][5][6]

Dig Insights' System 0 framework explains why the distinction matters. AI is beginning to mediate parts of the decision process, from research and shortlisting to recommendation and, in some cases, delegation.[7][8]

But the evidence does not support one universal ‘commercial influence’ number.

The studies use different prompts, platforms, time windows, classifications and outcomes. They should be read as separate layers of evidence, not pooled into a single result.

<h2>Start with the measurement ladder</h2>

The cleanest way to read the research is to separate four stages.

<div class='overflow-x-auto not-prose my-8'><table class='w-full min-w-[780px] border-collapse text-left font-sans text-sm'><thead><tr class='bg-[#15303C] text-[#F7FAF8]'><th class='px-4 py-3'>Stage</th><th class='px-4 py-3'>What is being measured</th><th class='px-4 py-3'>What it does not prove</th></tr></thead><tbody><tr class='border-b border-[#CAD7D3] bg-[#EAF1EF]'><td class='px-4 py-3 font-semibold'>Source exposure</td><td class='px-4 py-3'>A page is retrieved or displayed as a citation</td><td class='px-4 py-3'>The page caused the answer</td></tr><tr class='border-b border-[#CAD7D3]'><td class='px-4 py-3 font-semibold'>Brand association</td><td class='px-4 py-3'>The answer names a brand or connects it with an attribute</td><td class='px-4 py-3'>The brand was preferred</td></tr><tr class='border-b border-[#CAD7D3] bg-[#EAF1EF]'><td class='px-4 py-3 font-semibold'>Recommendation</td><td class='px-4 py-3'>The answer explicitly suggests or selects a brand</td><td class='px-4 py-3'>A person trusted or acted on it</td></tr><tr><td class='px-4 py-3 font-semibold'>Choice or delegation</td><td class='px-4 py-3'>A person relies on AI, approves an action or delegates a decision</td><td class='px-4 py-3'>The upstream source caused the outcome</td></tr></tbody></table></div>

Most of the new evidence sits in the first three stages. None of the studies follows the entire chain from page publication to citation, recommendation and completed consumer choice.

That is not a weakness to dismiss. It is a boundary to preserve.

<h2>First layer: commercial pages appear as sources</h2>

Peec AI analysed a fixed set of non-branded software-review prompts across ChatGPT, Google AI Mode, Perplexity, Microsoft Copilot, Gemini and Google AI Overviews. Across 12 calendar weeks from December 2025 through February 2026, it tracked 232,000 citations and 13,000 unique listicles.[1]

Peec reported that self-promotional listicles accounted for approximately 10% to 11% of the citations in its selected sample. Its reported platform averages were 3.6% for ChatGPT, 10.3% for Google AI Mode and 10.4% for Perplexity.

Those figures do not describe all AI search. The study focused on non-branded software-review prompts and selected the top 1,000 listicles per platform per week by retrieval count. It did not publish the complete prompt set, exact self-promotional numerator or platform-specific citation denominators.

<blockquote>In Peec AI's bounded software-review sample, self-promotional listicles remained present among tracked citations.</blockquote>

A separate DataPulse/GetCited research programme examined templated ‘best’, ‘which’ and ‘top’ comparison prompts across ChatGPT, Google AI Overviews and Perplexity.[2][3]

Its September English presentation reported that 25.4% of 106,758 matched displayed citations pointed to pages where the scanner detected at least one disclosed commercial-interest marker. Among 8,536 English answers with at least one matched source, 86% contained at least one marked page.

The marker mix was led by affiliate disclosure at 21.4% of citations. Partner content accounted for 2.9%, named commerce sections for 2.0% and advertising labels for 1.8%, which the publisher described as an upper bound. These categories overlap and cannot be added.

The study did not show that 25.4% of recommendations were paid. It did not test whether a page's commercial status caused its inclusion. It detected disclosure markers on cited pages.

PPC Land's methodological review reached the same essential boundary: citation observation is not recommendation formation.[4]

<h2>Second layer: citation does not guarantee brand association</h2>

Ahrefs provides a useful warning against treating a citation as an endorsement.[5]

It published 34 self-promotional pages across five domains for two Ahrefs brands, then tracked custom prompts in ChatGPT, Gemini, Perplexity and Copilot from 7 February to 31 May 2026. The study covered 9,886 answers.

Ahrefs Evolve appeared in 72 assistant-query slots where it had been absent during the six-day baseline. Ahrefs reported that one of its pages was cited in 82% of those new appearances.

That is an interesting temporal association. It is not a randomised experiment. The study had no matched untreated brand, no contemporaneous control group and no published adjustment for differences in prior awareness, query fit or third-party coverage.

The same study also found that a cited page did not reliably carry the promoted brand into the answer. Among answers citing a conference-promoting page, 43% omitted Ahrefs Evolve. Among answers citing an Ahrefs-tool page, 11% omitted Ahrefs.

The source was present. The promoted brand was not always carried through.

That distinction is fundamental. A page may contribute background facts, category definitions or competitor context without transferring its self-promotional claim into the final answer.

<h2>Third layer: some cited listicles coincide with more recommendations</h2>

Scrunch studied a narrower question: when an existing self-promotional listicle was cited for the same monitored question, was its author brand more likely to be recommended?[6]

Its public study used 818 self-ranking treatment pages, 1,033 neutral comparison controls and 5,434 treatment page-response pairs across seven AI surfaces.

Scrunch reported an author-brand recommendation rate of about 7% when the listicle was cited versus about 4% in matched answers where it was not cited. It also reported that the author brand was named about 39% of the time when the listicle was cited versus approximately 19% when it was not.

The recommendation result received Scrunch's temporal falsification check. The mention result did not.

The article also contains an important counterweight. In 63.9% of answers citing a self-promotional listicle, the system recommended nobody. A competitor was recommended instead of the author brand in 24.3% of listicle-citing answers.

Scrunch's design is stronger than a simple before-and-after comparison, but it is still observational. It did not randomise publication, ranking language, competitor inclusion or outbound links. Its capped causal analysis covered roughly 6% of the eligible population, and the exact cell counts were not disclosed.

The safe conclusion is not ‘crown yourself number one and AI will believe you’.

<blockquote>In Scrunch's monitored-question sample, citation of an existing self-promotional listicle was associated with a higher author-brand recommendation rate, but citation often produced no recommendation or a recommendation for a competitor.</blockquote>

<h2>The missing link is the one marketers care about most</h2>

The studies establish that commercial content can enter the evidence environment around an AI answer.

They provide some bounded evidence about when a brand is named or recommended.

They do not establish a general causal chain from commercial page to consumer choice.

That missing link becomes more important as AI moves further into the decision process. In the inaugural episode of <em>Between Two Joels</em>, Dig Insights framed System 0 as a way to think about AI mediating choices, from research and consideration to brand recommendation.[7]

The episode reported that 65% of surveyed consumers planned to use AI in some part of a near-future buying decision. The public transcript does not disclose the sample size, geography, field dates, sampling or weighting. The figure should therefore be treated as a publisher-reported stated preference, not a population estimate.

Dig's separate System 0 explainer reports that 65% prefer to use AI in some way when making a decision and 50% like AI to suggest purchase options.[8] It likewise does not publish the underlying sample or method.

System 0 is useful here as a decision lens. It does not validate the commercial-content studies or prove that citations change consumer behaviour.

<h2>Machine Brand Equity is the construct to measure</h2>

AEO Updates introduced <strong>Machine Brand Equity</strong> in the <a href='/articles/who-owns-ai-decision-layer-2030-scenarios'>decision-layer scenario framework</a>.[9]

<blockquote><strong>Machine Brand Equity is the strength of a brand's association with a relevant need, audience or decision territory inside machine-mediated choice.</strong></blockquote>

This is not the same as citation volume.

A brand can be cited frequently yet remain weakly associated with the attributes that matter in a decision. A competitor can be mentioned less often overall but appear consistently when a high-value need, audience or decision territory is expressed.

The commercial-content studies point toward a practical operational model.

Instead of asking only whether a brand appears, a study can pre-specify:

<ul><li>the decision territory, such as enterprise security, easy implementation or best value;</li><li>the competing brands eligible for comparison;</li><li>the attributes relevant to that decision;</li><li>the prompt families and phrasing variants;</li><li>the engines, modes, dates and rerun cadence;</li><li>the distinct outcomes to record.</li></ul>

Those outcomes should remain separate:

<div class='overflow-x-auto not-prose my-8'><table class='w-full min-w-[720px] border-collapse text-left font-sans text-sm'><thead><tr class='bg-[#15303C] text-[#F7FAF8]'><th class='px-4 py-3'>Measure</th><th class='px-4 py-3'>Denominator</th></tr></thead><tbody><tr class='border-b border-[#CAD7D3] bg-[#EAF1EF]'><td class='px-4 py-3 font-semibold'>Association rate</td><td class='px-4 py-3'>Eligible responses that connect the brand with the pre-specified attribute</td></tr><tr class='border-b border-[#CAD7D3]'><td class='px-4 py-3 font-semibold'>Consideration rate</td><td class='px-4 py-3'>Eligible responses that include the brand in a relevant shortlist</td></tr><tr class='border-b border-[#CAD7D3] bg-[#EAF1EF]'><td class='px-4 py-3 font-semibold'>Recommendation rate</td><td class='px-4 py-3'>Eligible responses that explicitly recommend the brand</td></tr><tr class='border-b border-[#CAD7D3]'><td class='px-4 py-3 font-semibold'>Evidence-support rate</td><td class='px-4 py-3'>Responses making the association that cite evidence supporting it</td></tr><tr><td class='px-4 py-3 font-semibold'>Source exposure</td><td class='px-4 py-3'>Eligible responses that retrieve or cite a relevant page</td></tr></tbody></table></div>

This is an operational research profile, not a validated composite score.

A single index would hide too much. It could reward citation volume even when the brand is not mentioned. It could count mentions that carry no recommendation. It could ignore whether the cited evidence actually supports the attribute being claimed.

<h2>Commercial content belongs in the evidence audit, not the shortcut playbook</h2>

The practical lesson is not to publish more self-ranked listicles.

It is to understand the information environment around the decision a brand wants to influence.

A useful audit should ask:

<ol><li>Which pages are being retrieved and cited for the relevant prompt family?</li><li>Which of those pages contain affiliate, partner, advertising or self-promotional signals?</li><li>Is the brand named when its own or third-party evidence appears?</li><li>Which attributes are attached to the brand?</li><li>Is the brand included in the consideration set?</li><li>Is it explicitly recommended?</li><li>Which sources substantiate the recommendation?</li><li>Does the result survive paraphrases, engines, modes and repeated runs?</li></ol>

This is where first-party and third-party evidence come together.

A brand's own pages can provide accurate product facts, claims, documentation and comparison logic. Independent sources can provide corroboration, context and credibility. AI systems may retrieve both, but the studies reviewed here show that appearance alone does not tell us how the evidence will be used.

<h2>What marketers should test next</h2>

<h3>Longitudinal tracking</h3>

Monitor the same prompt families, attributes and brands over time. Separate model updates, retrieval changes and content changes.

<h3>Controlled content interventions</h3>

Change one content element at a time, such as a substantiated attribute claim, comparison structure or source citation. Pre-register the hypothesis and compare treatment pages with matched controls.

<h3>Category-level benchmarks</h3>

Measure how Machine Brand Equity differs across categories, audience segments and decision stakes without pooling incompatible prompt families.

<h3>Choice experiments</h3>

Test whether AI-generated recommendations change human consideration and choice under controlled conditions. This is the downstream link the current commercial-content studies do not provide.

<h3>Provenance-aware analysis</h3>

Record which source supports each association or recommendation. Distinguish first-party claims, independent evidence, affiliate content, sponsored content and undetermined commercial status.

<h2>AEO Updates Takeaway</h2>

Commercial and self-promotional content can surface in AI answers. That is now supported across several bounded publisher datasets.

But the evidence does not support a universal recommendation effect.

Peec AI and DataPulse/GetCited primarily measure source exposure. Ahrefs shows that a cited promotional page does not always carry its brand into the answer. Scrunch provides bounded evidence that citation can coincide with more author-brand recommendations in a matched monitored-question sample, while also showing that most cited-listicle answers recommended nobody.

System 0 explains why the path matters. AI is beginning to mediate research, shortlisting and recommendation. The consumer-choice effect is still largely unmeasured.

That makes Machine Brand Equity the more useful strategic question:

<blockquote><strong>When a relevant need or decision territory is expressed, is the brand associated with the right attributes, included in the right consideration set and supported by credible evidence?</strong></blockquote>

The opportunity is real.

So is the need for measurement discipline.

<h3>References</h3>

[1] <a href='https://peec.ai/blog/self-promotional-listicles-analysis-from-232k-citations' target='_blank' rel='noopener noreferrer'>Tom Wells, Self-promotional listicles analysis: Data from 232,000 citations</a>, Peec AI, 3 March 2026.

[2] <a href='https://www.datapulse.de/versteckte-werbung-ki/' target='_blank' rel='noopener noreferrer'>Nicolas Caramella and Maria Fernandez, Die versteckte Werbung der KI</a>, DataPulse Research, 26 May 2026.

[3] <a href='https://www.getcited.media/studies/ai-commercial-sources/' target='_blank' rel='noopener noreferrer'>Maria Fernandez, One in four sources AI shows you for a best product question has a commercial interest</a>, GetCited, 7 September 2026.

[4] <a href='https://ppc.land/86-of-ai-best-product-answers-cite-a-commercial-source-getcited-finds/' target='_blank' rel='noopener noreferrer'>Luís Rijo, 86% of AI best product answers cite a commercial source, GetCited finds</a>, PPC Land, 27 September 2026.

[5] <a href='https://ahrefs.com/blog/self-promotional-content-ai-seo-experiment/' target='_blank' rel='noopener noreferrer'>Mateusz Makosiewicz, Self-Promotional Content Works, Until It Backfires: AI SEO Experiment</a>, Ahrefs, 6 July 2026.

[6] <a href='https://scrunch.com/blog/what-happens-when-you-crown-yourself-number-1-listicles-impact-on-ai-answers' target='_blank' rel='noopener noreferrer'>Niharika Sharma, What happens when you crown yourself #1</a>, Scrunch, 24 September 2026.

[7] <a href='https://share.transistor.fm/s/f1a9a4fa/transcript' target='_blank' rel='noopener noreferrer'>Dig Insights, 01: System 0</a>, Between Two Joels transcript, 26 May 2026.

[8] <a href='https://diginsights.com/resources/system-0/' target='_blank' rel='noopener noreferrer'>Dig Insights, Introducing: System 0</a>, publication date not disclosed.

[9] <a href='/articles/who-owns-ai-decision-layer-2030-scenarios'>Ian Ash, Who will own the decision layer?</a>, AEO Updates, 23 September 2026.

<p class='text-sm italic'>These studies are not pooled because their units, classifications, selection rules and designs differ.</p>