LLM Product Recommendations: Why Great PR Still Fails to Win

LLM product recommendations are becoming a new, high-intent discovery channel, but a familiar marketing assumption is already breaking down: great press coverage does not automatically make a product recommendable in ChatGPT, Gemini, Claude, Perplexity, or Google’s AI experiences.

Amanda Milligan’s analysis of the Chirp Contour massage table makes that gap unusually clear. In her video case study, Milligan found that a product with strong reviews and mentions from recognizable publications was still routinely absent when she asked large language models for the best at-home massage table. The issue was not simply that the product lacked publicity. It was that the signals surrounding the product did not give AI systems a sufficiently clear answer to three questions: Who is this for? What is it distinctly better at? And which sources already frame it as a top choice? (youtube.com)

That distinction matters for founders, ecommerce teams, SaaS marketers, and digital PR leaders. AI-driven product discovery is not a replacement for search, SEO, reviews, merchandising, or public relations. It is a layer on top of all of them—one that compresses a messy web into a few recommendations. If your positioning is vague, your product information is incomplete, or the sources consulted by the system do not include you, the model has little reason to put your brand in the final answer.

The Chirp Contour Case Study: Visibility Is Not Recommendation

Milligan’s original analysis centers on Chirp Contour, an at-home massage and decompression table. Her starting premise is intuitive: Chirp had earned coverage from outlets such as CNN, Lifehacker, Apartment Therapy, PopSugar, and ELLE, so one might expect the product to show up prominently in AI-generated recommendations.

Yet, when asking multiple LLMs for the best at-home massage table, the Contour was missing or appeared inconsistently. In the one context where it was surfaced, the recommendation was tied to at-home spinal decompression—a more specific use case than the brand’s product page appeared to emphasize clearly. Milligan also observed that, when asking how the Contour differed from alternatives, an AI system leaned on a third-party golf publication rather than Chirp’s own materials to explain the product’s positioning. (youtube.com)

The lesson is not that a massage table should be marketed through golf sites, nor that mainstream editorial coverage has stopped mattering. The lesson is more precise: broad awareness and positive sentiment are not identical to recommendation readiness.

A model asked for “the best” product is performing a compressed decision task. It is trying to infer category fit, compare trade-offs, identify credible evidence, and provide a shortlist that appears useful for a particular buyer. In that setting, a brand mention without explicit context is a weak signal. A detailed product comparison, ranked buying guide, specialist review, structured product feed, or first-party page that names its ideal use cases can be much easier for an AI system to turn into a recommendation.

Why LLM Product Recommendations Work Differently From Brand Mentions

Traditional PR is often evaluated through reach: the authority of a publication, estimated audience size, referral traffic, links, social engagement, and the credibility conferred by an editorial mention. Those outcomes remain valuable. But LLM product recommendations introduce a narrower question: can the product be selected and justified for a specific prompt?

An AI answer is usually not a popularity contest. It is closer to an instant buyer’s guide assembled from available information. The system may look for product attributes, reviews, retailer information, editorial comparisons, user-generated discussions, price and availability data, and sources that directly answer the buyer’s question.

Mention coverage can be too broad to help

Consider the difference between these two mentions:

• “Chirp released a sleek new wellness product for home use.”
• “Chirp Contour is an at-home decompression table designed for people who want guided back, neck, and shoulder relief without booking regular bodywork appointments; its key distinction is a contoured design intended to support positioning during use.”

The first may be positive and high-profile, but it tells a retrieval or reasoning system very little about when to choose the product. The second provides category, audience, job-to-be-done, alternatives, and differentiators. It is recommendation material.

This does not mean marketers should write for robots or fill pages with awkward “best for” claims. Google’s guidance for generative search continues to stress the same foundation as conventional SEO: create helpful, original, people-first content, ensure crawlers can access it, and make the page useful for searchers with increasingly specific questions. (developers.google.com)

The practical shift is to make the information buyers genuinely need explicit instead of assuming readers, journalists, search engines, and models will infer it from brand aesthetics or scattered coverage.

Recommendation prompts are inherently conditional

Most commercially meaningful prompts contain hidden qualifiers. A user may type “best email service,” but mean one of several things:

1. Best for a developer who needs a straightforward API.
2. Best for a startup sending receipts and password resets.
3. Best for a marketer who wants campaigns, templates, and segmentation.
4. Best for a regulated business with deliverability and compliance requirements.
5. Best for a team migrating away from a more expensive provider.

The same is true for mattresses, accounting software, project-management tools, exercise equipment, supplements, and consumer electronics. A model cannot credibly recommend one product for every buyer. Brands that explain the buyer, situation, constraint, and trade-off are far more legible than brands that only claim to be “premium,” “innovative,” “built for everyone,” or “the ultimate solution.”

Pitfall One: Your Product Page Does Not Name Concrete Use Cases

The first weakness Milligan identified was vague product messaging. The Chirp Contour page, in her view, left too much of the product’s primary value implicit. That matters because LLMs commonly segment recommendations by use case: best for small spaces, best for beginners, best for serious athletes, best for recovery, best under a budget threshold, best for portability, and so on. (youtube.com)

A product page cannot assume that a model will combine a lifestyle image, a feature list, and a review snippet into a precise explanation of why someone should buy. It may do so occasionally, but brands should not build growth plans around that hope.

Translate features into decision-ready language

A feature tells a buyer what the product has. A use case tells a buyer when it matters. A differentiated benefit explains why that feature beats an alternative.

For example:

The right copy depends on evidence. Do not invent medical, technical, financial, sustainability, or performance claims simply to sound more specific. But if a product has substantiated strengths, surface them in plain language, near the top of the page, in headings, in comparison sections, and in FAQ content.

Build an ICP-to-query map before rewriting copy

The most productive starting point is not a list of keywords. It is a map from ideal customer profile to real decision query.

For each priority audience, document:

• Their situation and trigger event.
• The job they are trying to complete.
• The alternatives they are considering, including doing nothing.
• Their non-negotiable constraints: price, size, skill level, integrations, location, time, risk, or compliance.
• The product capability that changes the decision.
• The proof needed to believe the claim.
• The exact language they use when asking a search engine, an AI assistant, a peer, or a reviewer.

A B2B example might look like this: “A seed-stage SaaS company needs reliable transactional email for account verification and password resets, wants developer-friendly integration, and does not want enterprise contracts or a sprawling marketing suite.” That is much more usable than “companies that need email.”

Once the map exists, turn it into page architecture. Create a clear category statement, use-case sections, comparison pages where appropriate, implementation documentation, pricing context, evidence, and support content that answers the natural follow-up questions a buyer will ask.

Pitfall Two: The Product Is Distinctive, but the Narrative Is Not

Milligan’s second point is especially important for products that look or function differently from their category peers. The Chirp Contour may be visibly unlike a standard portable massage table, but visual difference alone does not guarantee explanatory difference. If the manufacturer does not define the category and contrast it with familiar alternatives, other sites—or an LLM’s own inference—will fill the gap. (youtube.com)

That is risky because third parties may describe the product accurately but incompletely. They may lead with an angle that serves their audience rather than yours. A golf publication might focus on recovery; a design publication might focus on appearance; a shopping roundup might focus on price. None of those frames is necessarily wrong. But a brand should not outsource its core positioning to whichever article a model happens to retrieve.

Differentiation has to be comparative

Many positioning statements fail because they describe an aspiration instead of a contrast. “Simple,” “smart,” “modern,” “trusted,” and “all-in-one” are too elastic to guide a recommendation.

A useful differentiation statement has four pieces:

1. The category: What kind of product is this?
2. The intended buyer: Who gains the most from it?
3. The distinct mechanism or capability: What does it do differently?
4. The trade-off: What is it not optimized for?

For example: “This is a developer-first transactional email platform for product teams that need fast API integration and dependable application messaging; it is not intended to replace a full marketing automation suite.” That statement helps the right buyer self-select, helps a reviewer compare accurately, and gives an AI system a defensible way to recommend—or exclude—the product.

Say what you are not

Marketers often resist trade-offs because they fear narrowing the addressable market. In practice, trade-offs can improve credibility and conversion. A product that is “best for everyone” is difficult to believe and difficult to recommend.

A comparison section can clarify:

• When a customer should choose your product.
• When a more specialized or more full-featured alternative is preferable.
• Which features are intentionally absent.
• What operational burden the buyer avoids.
• How pricing, implementation time, control, or support differs.

This is not an invitation to publish attack pages or make unsubstantiated competitor claims. It is an invitation to make a real buying decision easier. Clear trade-offs reduce poor-fit leads, improve customer expectations, and create language that reviewers and AI systems can reuse accurately.

Pitfall Three: You Earned Coverage, but Not in the Citation Core

The most strategic insight in Milligan’s video is her description of a brand’s “citation core.” The phrase refers to the small set of pages and publishers that repeatedly show up as sources when LLMs answer the category queries you care about. In the massage-table example, the recommendation sources she found were ranked listicles, not simply the respected outlets that had reviewed Chirp positively. (youtube.com)

This is a crucial distinction. A media hit can be excellent for awareness, authority, links, and sales. But if it does not appear in the sources used for prompts like “best at-home massage table,” it may not move the recommendation result for that prompt.

Do not assume one AI system has one stable source set

Citation patterns can change by model, prompt formulation, user location, browsing settings, freshness requirements, and product availability. Amanda Milligan has also cautioned against overreacting to short-term LLM-source volatility, noting that source composition can change substantially over a short period. (backlinko.com)

That is why the citation core should be treated as a research input rather than a static target list. The goal is not to manipulate a single answer on a single day. The goal is to understand what evidence ecosystem shapes the category and to earn durable representation within it.

How to identify your citation core

Run a recurring audit using 15 to 30 high-intent prompts. Include both broad and constrained queries, such as:

• “Best [category] for [use case].”
• “[Product category] for [specific constraint].”
• “What is the difference between [your category] and [alternative category]?”
• “Best [category] under [$X].”
• “Which [category] is easiest to set up?”
• “What should I buy if I need [outcome] but not [common downside]?”

Use the AI systems your audience actually uses, and record the response date, prompt, brands named, URLs cited, source type, factual claims, and whether your own product page is included. Then categorize the recurring sources:

• Specialist editorial reviews.
• Ranked “best” lists and buying guides.
• Retailer or marketplace category pages.
• Manufacturer product pages.
• Comparison sites.
• Forums and community discussion threads.
• Industry publications.
• Video reviews and transcripts.

Patterns will emerge quickly. If nearly every answer cites a handful of specialist buying guides, pitching general news desks alone will not solve the problem. If first-party pages are appearing but your product is not, the issue may be positioning, crawlability, structured data, or insufficient proof. If retailer data is consistently present, catalog accuracy and availability may be the practical bottleneck.

The New Layer: Product Feeds and Structured Commerce Data

The Chirp Contour example is fundamentally about content and third-party narrative, but product discovery has moved further since that case study. In 2025 and 2026, major platforms expanded commerce-specific ways for merchants to provide structured product information.

Google recommends Product and Merchant Listing structured data so it can better understand product pages and potentially show richer shopping experiences, including price, availability, reviews, shipping, returns, and variants. Google also says that Merchant Center product data can improve its understanding of a product and is required for some shopping surfaces. (developers.google.com)

OpenAI has likewise introduced shopping research and richer product discovery features in ChatGPT. Its Agentic Commerce documentation says product feeds are designed to provide accurate catalog data for discovery, including current pricing, availability, and seller context; feed quality can improve relevance and reduce purchase friction. (openai.com)

Structured data does not replace positioning

A perfect feed cannot compensate for a product story that no one can understand. It may help a platform know that an item exists, costs $299, is in stock, and has a particular color or size. It does not, by itself, establish why the item should be selected for a particular person.

Think of modern product visibility as four connected layers:

1. Product truth: Accurate titles, features, specifications, variants, price, availability, shipping, returns, and images.
2. Product meaning: Clear use cases, buyer fit, differentiation, limitations, and comparison language.
3. Independent evidence: Reviews, tests, expert analysis, customer proof, editorial lists, and credible mentions.
4. Technical accessibility: Crawlable pages, consistent entity information, structured data, feeds, and no avoidable indexing barriers.

When one layer is missing, recommendations become fragile. A compelling story without reliable data can lead to stale or inaccurate results. Great data without differentiation leads to commodity comparison. Reviews without a clear first-party narrative leave others to define the category. And strong pages that cannot be crawled or interpreted are effectively invisible.

A Practical Operating System for AI Recommendation Visibility

The useful response is not to chase every AI answer or flood the web with synthetic pages. It is to establish an operating system that improves the quality, consistency, and discoverability of the information buyers need.

Step 1: Define your recommendation jobs

Start with five to ten decisions you want your brand to win. Avoid vanity prompts such as “What is the best [huge category]?” unless you genuinely have broad category leadership.

Instead, identify decisions that align with product-market fit. A cybersecurity tool might target “best endpoint protection for a 50-person remote company.” A creator tool might target “best video captioning app for short-form content.” An email platform might target “best transactional email API for an early-stage SaaS.”

These queries become the backbone of research, content, PR, and product education.

Step 2: Audit first-party pages for ambiguity

Read your product page as if you knew nothing about the brand. Can a buyer answer the following in less than two minutes?

• What category is this product in?
• What problem does it solve?
• Who is the strongest fit?
• What result should a buyer reasonably expect?
• How does it work at a high level?
• What is different about it versus the closest alternatives?
• What are the relevant constraints or trade-offs?
• How much does it cost, and what changes by plan, package, or variant?
• What proof supports the claims?

If the answer requires watching a video, opening hidden accordions, interpreting brand language, or visiting several unrelated pages, simplify the path. Important decision information should be easy for people and crawlers to find in visible HTML.

Step 3: Create proof assets, not just promotional assets

A glossy launch announcement is rarely enough. Build assets that support evaluation:

• A comparison guide with fair criteria.
• A transparent methodology page.
• A technical specification or implementation guide.
• Customer stories organized by use case.
• Original data or testing where appropriate.
• Expert commentary that explains limitations as well as strengths.
• A product FAQ that addresses setup, safety, compatibility, maintenance, support, and returns.

This approach serves journalists and reviewers as well as users. It gives third parties better material to cite and reduces the chance that your product is described through generic boilerplate.

Step 4: Pitch for decision relevance

Do not stop pitching high-authority publications. Instead, evaluate prospective coverage through a second lens: does the outlet publish content that helps people decide what to buy?

A strong target may be a specialist publication with a trusted annual buying guide, a reviewer known for detailed methodology, or a category newsletter whose readers actively compare products. The ideal placement is not merely “a mention.” It is accurate inclusion in a relevant, maintained, high-quality resource where your product is clearly matched to a buyer need.

Step 5: Validate product data continuously

For ecommerce brands, check product structured data, Merchant Center feeds, availability, pricing, variants, shipping, and return information. For software, keep feature pages, documentation, plan limits, changelogs, integrations, and pricing current.

This is unglamorous work, but recommendation systems cannot reliably surface information that is contradictory or outdated. Google specifically notes that richer product experiences rely on understandable product information, while OpenAI’s product-feed guidance emphasizes current catalog details for discovery relevance. (developers.google.com)

Step 6: Measure beyond mentions

Track traditional PR and SEO metrics, but add recommendation-specific measures:

• Share of voice across a defined prompt set.
• Frequency of first-party citations.
• Frequency of inclusion in relevant third-party buying guides.
• Accuracy of product attributes in AI answers.
• Presence in constrained use-case prompts.
• Referral traffic and conversions from AI platforms, where measurable.
• Changes in branded and category search demand after major coverage.

Avoid treating every model response as ground truth. AI outputs can be incomplete, stale, or wrong. The more useful metric is directional: are the systems increasingly able to describe your product accurately and include it when the buyer scenario is genuinely a fit?

What Marketers Should Not Do

AI visibility pressure is creating predictable bad habits. The tactics below may waste resources, damage trust, or conflict with platform quality guidance.

Do not manufacture “best of” pages at scale

Publishing dozens of thin listicles that name your own product is unlikely to create durable authority. Google’s guidance warns against using generative AI to produce large volumes of low-value content and reiterates that scaled content abuse can violate spam policies. (developers.google.com)

If you publish comparison or recommendation content, make it genuinely useful: disclose your perspective, explain criteria, include meaningful alternatives, update facts, and add evidence a buyer cannot get from a generic template.

Do not confuse visibility with endorsement

A model mentioning your brand does not mean it understands the product or recommends it for the right reason. It may pull a stale price, attach the wrong use case, describe a discontinued feature, or rank the product based on an irrelevant article.

Accuracy is the first milestone. Recommendation share is the second. Conversion quality is the third.

Do not copy competitors’ language blindly

If every company in a category calls itself “AI-powered,” “seamless,” and “enterprise-grade,” repeating those labels will not make your product more distinguishable. Interview customers, review sales-call objections, study support tickets, analyze competitor gaps, and identify concrete language that reflects the real choice being made.

The Bigger Strategic Change: PR, SEO, Product Marketing, and Merchandising Are Converging

The Chirp Contour example is useful because it exposes an organizational problem, not just a content problem. In many companies, PR owns earned media, SEO owns search traffic, product marketing owns messaging, ecommerce owns the catalog, support owns FAQs, and engineering owns structured data. AI recommendations cut across all of those disciplines.

A buyer-facing answer generated by an LLM may depend on work from every one of those teams. The product needs accurate specs. The positioning needs a sharp use case. Independent reviewers need a credible story and proof. The site needs technical accessibility. The feed needs current price and availability. The business needs a process for finding and correcting inaccurate information.

That is why “generative engine optimization” should not be a separate keyword-stuffing initiative. It is better understood as recommendation readiness: the capability to make a product easy to understand, easy to compare, easy to verify, and easy to surface in the places buyers now ask questions.

Conclusion: Be Easier to Recommend, Not Merely Easier to Find

Amanda Milligan’s Chirp Contour case study has a straightforward message: excellent coverage and a good product can still fail to produce LLM visibility if the product’s use case, differentiation, and source ecosystem are misaligned with how AI systems form recommendations. (youtube.com)

For marketers, the response is not panic and not a rush to produce AI-flavored content. It is disciplined clarity. Define the buyer scenario. State why your product is different. Publish proof. Maintain accurate product data. Study the sources that repeatedly shape category recommendations. Earn inclusion where real buyers make comparisons.

The brands that win LLM product recommendations will not necessarily be the loudest brands. They will be the brands whose information makes a confident, well-supported recommendation easy.

FAQ

What are LLM product recommendations?

LLM product recommendations are product suggestions generated by AI assistants such as ChatGPT, Claude, Gemini, Perplexity, or AI-powered search experiences. They typically synthesize information from first-party sites, product data, reviews, editorial coverage, comparisons, and other web sources to answer a buyer’s prompt.

Why do positive media mentions not guarantee AI visibility?

A positive mention may create awareness without explaining a specific use case, competitive difference, or buyer fit. AI systems often need sources that directly support a recommendation, such as buying guides, detailed reviews, product comparisons, and clear first-party product pages.

What is a citation core in AI marketing?

A citation core is the recurring set of sources that appear when AI systems answer the category prompts most important to your business. It is a practical research concept popularized in Amanda Milligan’s work, not a fixed platform metric. Audit it regularly because sources and outputs can change. (youtube.com)

Does structured data improve LLM product recommendations?

Structured data does not guarantee inclusion or top placement, but it can help platforms understand product attributes such as price, availability, variants, reviews, shipping, and returns. It should support—not replace—clear positioning, useful content, and independent evidence. (developers.google.com)

What is the first thing to fix if my brand is absent from AI recommendations?

Start by testing a focused set of high-intent prompts and documenting which sources are cited. Then check whether your own product page clearly states the category, ideal customer, specific use cases, proof, differentiators, and trade-offs. Fixing ambiguity on first-party pages is usually more controllable than trying to immediately earn new media coverage.