Getting into AI's "top 5": winning best-of and recommendation queries

"What is the best CRM for a five-person sales team?" "Top five accounting tools for freelancers?" Such questions do not ask an assistant to judge options the user already knows; they ask it to choose the options. A best-of query, or recommendation query, is a prompt that asks a generative engine to produce a shortlist for a category, need or segment, and the brands on that shortlist receive the buyer's attention before any website is visited. Winning such queries means being included, for the right reason, in a list the engine assembles from sources it trusts.

This differs from the head-to-head questions examined in our article on comparison queries, where the user names the contenders. Below we look at how engines build shortlists according to their documentation and published studies, what a brand can legitimately do to qualify, what the evidence says about self-promotional "best of" lists, and how to measure inclusion across the many ways people phrase one request.

How is a best-of query different from an "X vs Y" comparison?

In a comparison the candidate set is given, and the engine's task is framing: which option suits which need. In a best-of query the engine must first decide who belongs in the set, and that is where most brands are eliminated without trace. A brand can win every comparison it enters and still lose the category because it is rarely entered; the dynamics behind that pattern are discussed in whether AI is recommending your competitor.

Best-of queries also arrive in families. "Best CRM for a small team," "top 5 CRM tools for startups" and "which CRM should a small agency use" express overlapping intents with different constraints, and each variant can yield a different list. Treating the family, not the sentence, as the unit of analysis is the methodological core of what follows.

How do AI engines assemble a shortlist?

The documented mechanics are retrieval followed by synthesis. Google's guide to generative AI features describes grounding answers through its core Search ranking systems, together with a query fan-out in which the model issues related queries to gather more material. OpenAI says ChatGPT search rewrites the question into targeted queries for search providers, ranks results on multiple factors and does not guarantee placement; for shopping it adds that products are selected using structured metadata from first-party and third-party providers, that review summaries are generated from public websites and are not verified by OpenAI, and that product results are not ads. None of these companies publishes the weights that decide who makes the final list, and any claim to know them deserves suspicion.

Published studies describe what the retrieved material tends to be. In an Ahrefs analysis of 750 top-of-funnel ChatGPT prompts, published in December 2025, "best X" blog lists made up 43.8 percent of all source links classified by page type, and the author, Glen Allsopp, observed a correlation between a brand's high position in third-party lists and its appearance in answers while noting that the mechanism was still under study. A 2025 arXiv preprint by Chen, Wang, Chen and Koudas found a systematic bias in AI search toward earned media, meaning third-party authoritative sources, over brand-owned and social content; for queries about well-known brands, 93.5 percent of the sources ChatGPT cited in their sample were earned media. The paper also documented a big-brand tilt: in fifty unbranded prompts asking for the "most popular," "best" or "top" colas, major brands took 56.3 percent of ChatGPT's brand mentions and niche brands 12.3 percent.

What evidence does an engine need to put you on the list?

An engine can recommend you only for a need it finds you associated with. The first requirement is unambiguous category positioning: a sentence on your site, repeated in your profiles, stating what you are, for whom and for which job. The second is a "who it is for" statement that also says who it is not for, because segment queries such as "for a small team" are matched against stated fit, not against adjectives like "leading." Chen et al. call this supplying "justification attributes," the extractable reasons an engine can attach to a name.

Criteria-based pages turn positioning into evidence. A page that explains how to choose in your category, which criteria matter for which buyer and where your product fits gives the engine both the frame and your place in it; our guide to the best CRM for a small business follows such a criteria-first structure. Consistency completes the picture: if your pricing model, integrations or languages are described differently on your site, on review platforms and in roundups, the engine receives competing attributes and has less reason to name you.

Why do independent roundups and review platforms carry so much weight?

Because a shortlist is a synthesis, the engine needs corroboration it did not get from you, and editorial roundups, analyst lists, review platforms and community threads supply it. Engines differ in degree: in the Chen et al. sample Perplexity drew 23.8 percent of citations for well-known brands from social sources, far more than ChatGPT, which helps explain why AI cites Reddit and forums so often in buying contexts. The legitimate route into these sources is slow but durable: give editors accurate information, make honest reviews easy on the platforms your buyers read, correct factual errors and earn inclusion on merit. Our article on testimonials, reviews and case studies covers collecting that proof with consent.

Should you publish your own "best X" list? What the evidence says

Self-promotional listicles, in which a vendor ranks its own category and places itself first, had become a widespread tactic of generative engine optimization (GEO) by 2025, and the evidence on them is mixed. Allsopp found that in the software category, when a publisher's own brand was recommended, its self-promotional list appeared among the cited pages in more than a third of ChatGPT responses, and he advised agencies and SaaS companies to publish such lists. Lily Ray then tracked 100 B2B "best software" queries in Google AI Overviews at three dates between April and June 2026; across the 80 prompts that produced an Overview, 224 of 323 cited self-promotional listicles, 69 percent, did not come with a recommendation of their publisher. She also reported that organic traffic at dozens of sites relying heavily on such pages began falling sharply around January 20, 2026, and attributed this to an algorithmic adjustment by Google, an interpretation Google has not confirmed.

Two further studies refine this picture. In an experiment published in July 2026, Ahrefs placed 34 pages promoting its own conference and its own AI visibility product on five domains and analysed 9,886 answers from ChatGPT, Gemini, Perplexity and Copilot collected between February and May 2026. The new conference entered 72 prompt-and-engine combinations in which it had not been named before, mostly in answers citing one of the pages, whereas for the established product only 6 percent of such new mentions came from answers citing them; and 43 percent of answers citing a conference page never mentioned the conference. Scrunch, a visibility-monitoring vendor, studied 818 self-ranking pages against 1,033 neutral third-party comparisons across seven AI surfaces. Holding prompt, brand and assistant fixed, it estimated that the publisher's recommendation rate rose from about 4 to roughly 7 percent when its own listicle was cited; that comparison was run on a capped sample covering about 6 percent of the eligible population, and a placebo test attributed a quarter to a third of the gap to brands that were already rising. In 63.9 percent of responses citing such a listicle, the assistant recommended no one. The defensible reading is that citation and recommendation are different outcomes, as our article on mention versus citation explains.

What do Google's guidance and consumer law say about self-published rankings?

Google's reviews system evaluates first-party content offering recommendations or analysis, and its guidance on high-quality reviews applies directly to "best" pages: explain what sets something apart from competitors, discuss benefits and drawbacks based on original research, and, when calling something the best for a purpose, say why with first-hand evidence. Its guide to generative AI features adds that seeking inauthentic mentions "isn't as helpful as it might seem," because those features depend on core ranking and spam-blocking systems, and its spam policies, in what Google now calls its site reputation policy, give the example of a low-quality third-party "best casinos" page placed on a medical site to benefit from that site's established ranking signals.

Consumer law draws a harder line in some markets. The US Federal Trade Commission's rule on reviews and testimonials, announced in August 2024, prohibits a business from misrepresenting that a website it controls provides independent reviews or opinions about a product category that includes its own products. A vendor's category page is legitimate when it is open about authorship; it becomes a problem when it poses as an independent referee.

A place on an AI shortlist cannot be claimed; it is earned by being the option that independent sources and your own transparent criteria describe, in the same terms, as the right fit for a stated need.

Where does optimization end and manipulation begin?

Research confirms that content changes can shift generative visibility. The GEO study by Aggarwal and colleagues, presented at KDD 2024, built a benchmark of 10,000 queries, GEO-bench, and found that adding quotations, statistics and source citations improved visibility: the best method raised a position-adjusted word-count measure by about 41 percent over the unoptimized baseline, while keyword stuffing fell below that baseline on the same measure. The darker edge is documented too: Kumar and Lakkaraju (2024) showed that a crafted text sequence added to a product's information page could significantly raise the chance of that product, in a catalog of fictitious coffee machines, becoming an LLM's top recommendation, and Pfrommer and colleagues showed at EMNLP 2024 that prompt injections on product websites could promote low-ranked products, and that the attacks transferred to deployed conversational search engines such as Perplexity. The boundary is evidential: clarifying who you serve and citing real data help an engine describe you correctly, whereas hidden instructions, fabricated reviews and fake independence try to make it describe you incorrectly.

How do you measure inclusion and position across phrasing variants?

A single screenshot is not evidence. SparkToro and Gumshoe had 600 volunteers run 12 prompts through ChatGPT, Claude and Google's AI Overviews (or AI Mode when no Overview appeared) 2,961 times in late 2025 and found less than a 1 in 100 chance of getting the same list of brands twice and about 1 in 1,000 of getting it in the same order. They concluded that rank in a single answer means little, whereas visibility percentage across many prompts run repeatedly is a reasonable metric, and noted that real users almost never phrase the same intent alike.

Build a prompt family for each category intent, covering generic, segment, "top N," "alternatives" and language variants drawn from prompt research. Run each variant repeatedly, use inclusion rate as the primary metric, read position as a distribution, and compare against a fixed competitor set as in our guide to share of voice in AI search. Why answers move between runs is the subject of our article on volatile AI answers, and the local form of the problem is covered in local GEO.

Rocketly GEO Suite supports this workflow within stated limits. Its Competitor Analysis reports share of voice, mention rate, citation rate and rank, and offers brand names that recur in measured answers as candidates for review rather than adding them automatically. Question variants track an alternative wording separately and compare rates on the same engine over a 30-day window, which the product itself says is not a causal experiment. There is no dedicated best-of feature: prompts containing "best" map to a category-page suggestion in the action plan, and any content draft it prepares is reviewed and approved by a different person and never published automatically.

The shortlist decision table

Shortlist factor	Evidence the engine can find	Action	Measurement
Category association	Pages naming you within the category	One-sentence category definition on site and profiles	Inclusion rate on generic prompts
Segment fit	"Who it is for" statements, use cases	State ideal customer and non-fit	Inclusion on segment variants
Independent roundups	Editorial lists, analyst reports	Supply accurate facts; correct errors	Roundups engines cite for your prompts
Review platforms	Ratings and review text	Ask all customers for honest reviews	Volume, recency, platforms cited
Attribute consistency	Pricing, integrations, languages across sources	Align core attributes everywhere	Attributes in answers versus your facts
Own criteria page	Transparent category page	Disclose authorship, criteria, fair rivals	Cited, and whether citation leads to recommendation
Position	Order of names in the answer	Strengthen the factors above	Position distribution over repeated runs

To see how often your brand enters category shortlists in generative answers, start with Rocketly GEO Suite: Google AI Overview is measured weekly as standard, ChatGPT and Perplexity are in controlled rollout, and Gemini, Claude and Grok are not measured.

Frequently asked questions

Can I pay to be included in an AI shortlist?

OpenAI states that ChatGPT's product results are selected independently, are not ads and are not influenced by OpenAI partnerships, and that ads are kept separate from product results. None of the documentation reviewed for this article describes a way to buy a position inside a generated list.

Can my own category guide include my product?

Yes, if the page is open about who wrote it, explains its criteria, treats alternatives fairly and says when another option is the better choice. It should never present itself as an independent review site; in the United States the FTC rule expressly prohibits that misrepresentation.

How many prompt variants should I track per category?

We know of no established threshold, and the SparkToro authors list the number of prompts needed as an open question. Cover each distinct intent and constraint your buyers express, run every variant repeatedly, and add variants until the family's inclusion rate stops moving materially.