How Do AI Assistants Decide Which Companies to Recommend?

Recommendation is not citation

The distinction sounds pedantic and decides the whole question. When an assistant answers a factual question and footnotes the pages it drew on, it is citing sources. When someone asks for the best options in a category and it names five companies, it is recommending.

These are different outputs with different stakes. A citation credits a document; a recommendation places a company on a shortlist, which is closer to what a buyer of visibility actually wants and is the harder of the two to influence.

Almost all available documentation concerns the first. Google publishes what makes a page eligible to appear as a supporting link. Microsoft reports which of a publisher's pages were cited. OpenAI publishes which crawlers fetch content and how to control them. Every one of these describes documents being used as sources.

None of them describes how an entity comes to be named as an option. That is the question this report is about, and the honest headline is that no operator has answered it.

The confusion is understandable because both appear in the same answer. A response naming five vendors may also footnote the pages it drew on, and the footnotes are visible while the selection is not. A company can therefore watch its competitors being recommended without learning anything about why.

What do the operators publish about recommendations?

Nothing specific, and the silence is consistent across all three.

Google's AI features documentation addresses eligibility for pages, stating that to appear as a supporting link a page must be “indexed and eligible to be shown in Google Search with a snippet”, and that there are “no additional requirements…nor other special optimizations necessary”. That is a statement about documents. It says nothing about which businesses are named inside an answer.

Microsoft goes furthest on transparency and stops at the same boundary. Its AI Performance report tells a publisher which of their URLs were cited and what grounding queries retrieved them, while explicitly cautioning that the metrics reflect how often content is referenced without indicating “ranking, authority, or the role of any page within an individual answer”.

OpenAI publishes crawler identities and controls and nothing about selection at all. So a company asking why a competitor appears in an assistant's shortlist and it does not has no documented answer available from any of them.

What can reasonably be inferred?

One thing, and it should be labelled as inference rather than finding.

These systems answer by retrieving documents and synthesising from them. A question asking for the best options in a category will retrieve documents that discuss options in that category — comparison articles, roundups, review pages, buyer's guides, forum threads. Companies named in those documents are the raw material from which any shortlist can be assembled, because there is nothing else for the system to draw on.

That reasoning follows from how retrieval-based answering works and is consistent with the published evidence on citation. Research examined earlier in this series found encyclopaedic and reference sources dominating citations, which tells us something about what these systems reach for and nothing about how they rank vendors.

What cannot be inferred is anything quantitative or actionable at the level being sold. Nobody can say how many mentions matter, which publications carry weight, whether recency dominates, or how any of it differs between assistants. Those questions have no published answers, and a specific figure offered for any of them was produced by observation and guesswork.

There is a further asymmetry between the two acts. A citation attaches to a specific document that can be improved, replaced or corrected. A recommendation attaches to an entity, and there is no document to edit — which is why the ordinary levers of publishing apply so poorly to it.

Why is this particularly hard to influence honestly?

Because the documents that plausibly feed recommendations are precisely the documents a company cannot legitimately write about itself.

A comparison article listing the best vendors in a category derives whatever value it has from having been written by someone other than the vendors. The same structural point applies here as applies to notability, to visa evidence and to search ranking: the thing being sought is independent judgement, and arranging it destroys it.

This creates an obvious market and an obvious hazard. Services offering to place companies into roundups and listicles are selling insertion into exactly the substrate that plausibly matters — which, if the material is paid and undisclosed, produces the same problems documented throughout this series: an FTC disclosure obligation, exclusion from notability, and a durable public record of a purchased mention.

And the disclosure that makes such placement lawful is text on the page. A comparison article marked as sponsored is a document whose own content announces the ranking was paid for, which is a poor foundation for a claim of independent standing.

What about services promising AI recommendation?

Treat any specific promise with the scepticism the evidence warrants. No operator publishes how companies are selected for recommendation, so no vendor can hold a documented mechanism, and nobody can verify delivery.

The verification problem is worse here than for search rankings. A ranking can be checked by searching. An assistant's answer varies between users, sessions, phrasings and model versions, which means a screenshot proves that one response on one occasion named a company, and no more than that. Absence of a result is equally unfalsifiable.

Some of the underlying work is nonetheless legitimate and worth buying under its right name: accurate, indexable pages that answer real questions about a category; a consistent public description; genuine presence in places people and systems read. That is ordinary web and content work whose value does not depend on any claim about assistants.

What should be refused is the framing that turns unverifiable work into a deliverable. A promise to appear in AI recommendations is a promise about a process its operators decline to describe and that the buyer cannot audit.

Ask also what happens when the model version changes. Answers shift with releases that vendors do not control and are not told about in advance, so any position achieved is held at the sufferance of a deployment schedule.

What this means for anyone buying placement

Separate the two questions before spending anything. Being cited as a source is partly documented and partly achievable through ordinary means. Being recommended as a company is undocumented, and nobody selling it can show you how it works.

Ask what would count as delivery, and how it would be checked. If the answer is a screenshot of one conversation, the metric is not a metric. Assistants do not produce stable, reproducible outputs the way a search results page does.

Invest in the substrate honestly. If comparison content plausibly matters, then being genuinely worth including in it — a real product, real differentiation, customers willing to say so publicly — is the durable route, and it is the same asset that serves every other objective in this series.

Media Morpheus sells placement and search visibility work. We can produce indexed pages that answer specific questions, and we can reach real audiences. We cannot place a company into an AI assistant's recommendations, because no published mechanism exists for doing so and no one can verify it if they claim otherwise.

The category is young enough that the vocabulary is still being invented, which is usually when the least verifiable promises are easiest to sell.

Methodology and limitations

This report relies on published operator documentation from Google, Microsoft and OpenAI covering AI features, citation reporting and crawler controls, and on research into AI citation patterns examined earlier in this series. Five limits apply. First, the central finding is an absence: no operator publishes how companies are selected for inclusion in a recommendation, and the report reports that rather than filling it. Second, the discussion of comparison content as a plausible substrate is explicitly an inference from how retrieval-based answering works, not a published finding, and no measurement supports it. Third, no original testing of assistant outputs was conducted, because assistant responses vary by user, session, phrasing and model version, and small-scale probing produces anecdotes rather than data. Fourth, third-party studies claiming to measure AI recommendation share were not relied on, as those located do not disclose methodology sufficient to assess them. Fifth, operator documentation changes without notice and this describes the position at the date of publication.

Conclusion

There are two different questions hiding inside this one, and the market sells the answer to the easier while charging for the harder. Being cited as a source has partial public documentation: Google publishes eligibility conditions, Microsoft shows publishers which pages were cited, OpenAI publishes crawler controls. Being named as a recommended company has none. No operator describes how an entity reaches a shortlist, no method exists to verify a promise about it, and assistant outputs are too variable for a screenshot to constitute evidence. The only defensible inference is that these systems assemble shortlists from documents discussing the category, which points at comparison content that is worth appearing in and cannot honestly be bought. Everything more specific is invention.

Frequently Asked Questions

Sources

1. official Google Search Central. Documentation on AI features in Google Search, for eligibility conditions for AI Overviews and AI Mode.
2. official Microsoft. Introducing AI Performance in Bing Webmaster Tools, for citation reporting and its stated limitations.
3. official OpenAI. Crawler documentation for GPTBot, OAI-SearchBot and ChatGPT-User.