top of page

Cited, Not Chosen

Sep 8
14 min read

The finding: quoted in half the answers, named in none

We ran six prompts through ChatGPT, Perplexity, and Google AI Overviews over thirty days. The prompts were the questions a prospective client asks before they hire a brand agency: questions about US market readiness, international expansion, and why so many European brands fail in their first year in America.


The first finding was encouraging. On the prompt "How do we know if our brand is ready for the US market?", octonanonyc.com was cited in 52.6% of 19 responses. On "I'm planning to expand internationally, what makes a brand ready", cited in 38.9% of 18 responses. The site was informing the answer in more than a third of cases. That is not a visibility problem.


The second finding is the one that matters. Across all six prompts, Octonano was mentioned by name as an organization in 7.1% of responses to one prompt, and 0% of responses to the other five. Totalled across those six prompts: 118 responses, our content cited 27 times, our company named once.


The denominators are small, between 14 and 31 responses per prompt. The finding is not any single percentage. It is the consistent direction across six prompts: cited constantly, named almost never.


The commercial sting is in one row. On the prompt "Who can run a US market readiness assessment for our brand?", octonanonyc.com was cited in 13.3% of 15 responses and mentioned as a recommendation in 0%. That is a buying question. Our own writing on US market readiness informed the answer. Someone else was recommended.


The gap between those two numbers is what this article is about.


Cited, Named, Chosen: three different wins

The way most marketing teams talk about AI visibility, these three outcomes are treated as a single thing, or as a ladder where the lower rungs accumulate into the higher ones. They are not a ladder. They are three different questions, and answering one does not answer the others.

Outcome

What happened

What produced it

Cited

Your page informed the answer. A link may appear in the sources.

Extractable, well-structured, useful content on your own site

Named

Your organization appears in the answer as an entity

Entity clarity and consistency across independent sources

Chosen

You are the recommendation the reader acts on

Corroboration a stranger can verify without taking your word for it

The reason this matters is not semantic. It is operational. A team that conflates citation with recommendation will spend the next twelve months publishing more content, watching citation rates hold or rise, and reporting that AI visibility is improving. They are right, and the pipeline will not move, because they are optimizing for the first question while the recommendation is decided by the third.


What each outcome actually requires:

  • Cited is achievable through publishing. Structure your content so AI systems can extract a clean passage. Answer a specific question in the first sentence of a section. Use headings that match how people ask. This is documented, learnable, and within the control of any content team. If you want to understand whether AI systems can see you at all, that question is answered before this one.

  • Named requires that independent sources agree on who you are. Your site can say your name. It cannot make a third-party directory, a publication, or a knowledge graph entry say it. Entity consistency is built outside your own domain.

  • Chosen requires that a stranger, reading sources you did not write, would reach the same conclusion about your organization that you would reach yourself. That is a different kind of evidence entirely, and no amount of publishing produces it directly.


Why citation is the easy one

That is not a criticism. Citation being achievable through publishing is genuinely useful. It means a brand with good content and clean structure can be part of the conversation AI assistants are having with your prospective clients. That is not nothing.


The problem is the way citation has been marketed. Much of the AEO and GEO positioning in 2026 rests on the premise that citations are the metric to move, and that moving them is evidence of progress toward recommendation. The logic is appealing because it is partly true. A brand that is never cited is not going to be recommended. But the inference runs in only one direction.


Citation is necessary. It is not sufficient. And the gap between the two is where most AI-visibility programs stall.

What makes citation achievable is also what makes it insufficient for recommendation. AI systems select pages for citation based on documented criteria: whether the content directly answers the question, whether it is structured for extraction, whether the domain has the kind of authority signals that come from being indexed, crawled, and referenced. These are content and technical signals. They are all first-party or first-party-adjacent. A brand can optimize every one of them without a single independent source ever corroborating that the brand is trustworthy.


The platform split in our data reflects this. ChatGPT accounted for 46.5% of responses, Perplexity 35.3%, and Google AI Overviews 18.1%. Citation behavior varied across platforms. Recommendation behavior did not. Octonano was named in 0% of responses on five of six prompts regardless of which platform generated the answer. The content was sufficient for citation across all three. It was not sufficient for recommendation on any of them.


The Two Ledgers

This is our way of describing the distinction, not a mechanism any platform has published. We find it useful because it locates the problem precisely.

Every organization has two ledgers. AI assistants read them for different purposes.


The ledger you write. Your website, your blog, your service pages, your schema markup, your structured data. You control this completely. You can update it today, optimize it tomorrow, and publish fifty more pages next month. That control is exactly why it cannot settle the recommendation question. A source that can say anything about itself is useful for facts and useless for trust. AI systems use this ledger to understand what you do and to extract content for citation. They do not use it to decide whether you are the organization they should put their name behind.


The ledger others write. Reviews, directories, listings, editorial coverage, mentions in contexts you did not create and cannot edit. You do not control this ledger. You cannot update it on a deadline. You cannot optimize it in a sprint. That is precisely what makes it load-bearing. When an AI assistant is deciding whether to recommend an organization, it is looking for evidence that a stranger, with no stake in the outcome, reached a positive conclusion about that organization. The ledger others write is the only place that evidence exists.


Citation reads the first ledger. Recommendation reads the second.

The reason most AI-visibility work in 2026 is entirely first-ledger work is not that agencies are misguided. It is that the first ledger is the one that can be scoped, billed, and shipped in a sprint. The second ledger is slower, less controllable, and structurally outside the scope of a content program. That is not a reason to ignore it. It is a reason to name it, budget for it, and stop measuring it with citation metrics.


What assistants actually check before they recommend

No major platform publishes a recommendation formula. Google, OpenAI, and Perplexity document how their systems select sources for answers. None of them publishes a threshold of reviews, coverage, or entity consistency that makes a brand recommendable. That absence is itself informative: if there were a formula, this would be a checklist problem, and it would already be solved.


What the public documentation does describe is a consistent set of signal classes. When an AI assistant is asked to recommend a company, the signals it draws on are different from the signals it draws on when composing an informational answer.


Structured data and entity signals

AI systems that surface recommendations look for consistent entity signals across sources they did not write. Your name, your category, your location, and your service description should read the same way in your schema, in your Google Business Profile, in directory listings, and in any third-party coverage. Inconsistency is not a penalty in the way a ranking algorithm penalizes it. It is simply noise. A system that cannot resolve who you are with confidence will not put its name behind a recommendation of you.


Prominence, and why reviews are its most legible component

Google's public guidance on local ranking describes three factors: relevance, distance, and prominence. Prominence is the one that draws on the second ledger. It includes the volume and quality of reviews, the consistency of information across the web, and the presence of editorial coverage.


Reviews are the most important and most under-discussed component of prominence in AI recommendation conversations. They are not a standalone trigger. They are the most legible and most verifiable part of the prominence signal, and the one a company cannot write for itself. A review is a statement by a stranger, in a context the brand did not create, that a real transaction produced a real outcome. That is exactly the kind of evidence the recommendation question requires.

Most AEO and GEO commentary in 2026 does not connect reviews to AI recommendation at all. That is the gap this article exists to name.


The category nobody has won yet

Here is what the flat field in our data actually shows. Across the full set of prompts we track, not only the six above, Octonano holds 3 mentions and 1.6% mention share against 33 competitors. No brand among those 33 exceeds 1.6%. Five agencies sit tied at the top of that list, including ours.


There is no incumbent to displace. There is an empty chair.


This is not a deficit story. It is a land-grab story, and the distinction matters for how you allocate the next six months.


A deficit story says: we are behind, we need to catch up, we need more content. That framing produces more first-ledger work, which is the work that has already been done and has already produced the citation numbers without producing the recommendations.


A land-grab story says: no one has built the second-ledger presence this category requires, and the brand that does it first will hold a position that compounds. The signals that produce recommendation are not replicable in a sprint, which is what makes the position worth holding.


The category is not waiting to be disrupted. It is waiting to be claimed. The brands that understand the difference between citation and recommendation before their competitors do are the ones with the first-mover position. In our measurement, that position is currently vacant.


What to do about it

Before the practical direction: this article is itself a publication, and a sharp reader will notice the tension. We are arguing that you cannot publish your way to being recommended, and we are making that argument in a blog post. The tension is real and worth naming directly.


Content is how you earn the citation, which is necessary and not sufficient. This article is not a substitute for second-ledger work. It is an explanation of why second-ledger work is the next move. The system we build for search and AI visibility addresses both ledgers. This article addresses the one that is most often left off the brief.


First: measure the right thing

If your current AI-visibility report shows citation rates, add a measurement for brand mentions and recommendation share on the prompts that matter commercially. Citation and mention are different data points. Reporting one as evidence of the other is how the problem stays invisible.


Second: build the second ledger deliberately

Reviews, directories, listings, editorial coverage, and third-party mentions are not a byproduct of doing good work. They are a program. They require the same intentionality as a content calendar, and they operate on a longer timeline. Recency and velocity matter: a steady flow of fresh third-party signals is more useful than a historical archive of them.


Third: resolve your entity signals

Your name, category, location, and service description should be consistent across your site, your Business Profile, your schema, and every directory where you appear. Inconsistency does not penalize you algorithmically in a direct sense. It makes you harder to resolve, and a brand that is hard to resolve is a brand an AI assistant will not put its name behind.


The empty chair

The data we started with is uncomfortable, and it is also clarifying. A citation rate above 50% on a commercially relevant prompt, combined with a recommendation rate of 0%, is not evidence of failure. It is evidence of a very specific and very fixable gap.


The fix is not more content. The content is already working. The fix is building the evidence that citation cannot produce: the reviews, the coverage, the entity consistency, the third-party corroboration that makes a stranger's recommendation possible.


No brand in this category has done that work at the level AI assistants require. The chair is empty.


The question is not whether your brand can be recommended by AI assistants. The question is whether you move before the brand that is about to understand this does.


Where to start

The first move is measurement, and it is not the measurement most AI visibility reports show you. Run the buying questions your prospects actually ask, and record separately whether your domain is cited and whether your company is named. The gap between those two numbers is the size of the problem.


We run that as a diagnostic: your prompts, your category, your competitors, and the table this article opened with, built for your brand. It tells you which of the two ledgers is the constraint before you spend anything on fixing either.


If your content is being quoted and your name is not, talk to us.

Frequently Asked Questions:

How do you measure whether an AI assistant recommends your brand rather than merely citing it?

Measuring recommendation requires running the prompts a prospective buyer would actually use, not the informational prompts that produce citations. A buying prompt such as "who can help my brand with US market entry" is structurally different from "what makes a brand ready for the US market." Run both categories of prompt across ChatGPT, Perplexity, and Google AI Overviews, and record separately whether your domain appears as a cited source and whether your organization is named as a recommendation. Citation and mention are different data points in the response. Tracking only citation rates and reporting them as AI visibility progress is how the gap between the two stays invisible. The metric that matters commercially is mention share on buying prompts, not citation share on informational ones.

AI assistants draw on a range of sources outside the company's own website when forming recommendations. Publicly documented signals include review platforms, business directories, editorial coverage, and knowledge graph entries that describe the organization consistently. Google's guidance on local ranking identifies prominence as a key factor, which includes the volume and quality of reviews, the consistency of business information across the web, and the presence of coverage in independent sources. Perplexity and ChatGPT draw on indexed content from across the web, which means any third-party source that describes an organization accurately and positively contributes to the evidence base an AI assistant can draw on. No platform publishes a definitive list of approved sources. The principle is corroboration: independent agreement from sources the brand did not write.

No major platform publishes a threshold. Google does not state a minimum review count for local ranking, and neither OpenAI nor Perplexity documents a review volume requirement for recommendation eligibility. What the public guidance does describe is that recency and velocity matter alongside volume. A steady flow of recent reviews is more useful than a large historical archive with no new activity. The more useful question for most organizations is not how many reviews are needed but whether the reviews that exist are recent, specific, and distributed across the platforms an AI assistant is likely to draw on. An organization with consistent, recent, specific reviews across multiple independent platforms is in a stronger position than one with a high lifetime count concentrated on a single source.

An AI assistant recommending a company with no third-party coverage would be relying entirely on what the company has written about itself. That is structurally the same as a reference who is the subject of the reference. The recommendation question requires evidence a stranger can verify without taking the company's word for it. In the absence of third-party coverage, reviews, directory listings, or editorial mentions, an AI assistant has no independent corroboration to draw on. It may still cite the company's content in informational answers. It is unlikely to name the company as a recommendation in a buying context, because the evidence that a recommendation requires does not exist in the sources it can access.

There is no documented timeline, and any specific figure would be speculation. What is known is that second-ledger signals, reviews, directory listings, editorial coverage, and entity consistency, operate on a different timeline from content. A well-structured article can be indexed and cited within days. A review profile with sufficient recency and velocity, and editorial coverage that establishes an organization as a credible entity in its category, builds over months. The more useful framing is not duration but dependency: recommendation does not follow automatically from citation, regardless of how long citation continues. The two outcomes require different inputs, and the inputs for recommendation are not produced by a content program alone.

AI recommendation draws on signals that no single function controls. Marketing owns the first-ledger work: content, schema, structured data, and entity consistency across the company's own properties. PR and communications own the coverage, editorial mentions, and third-party placements that constitute second-ledger corroboration. Sales owns the relationships that generate reviews and client testimonials. The practical answer is that AI recommendation requires coordination across all three, with a single owner responsible for measurement. Without measurement that separates citation from recommendation, each function will optimize for the metric it controls and the gap will persist. Assigning ownership to marketing alone, which is the default in most organizations, produces citation programs that do not move recommendation share.

There is no fixed cost because the inputs vary significantly by organization. Review generation requires a systematic process for asking satisfied clients to leave reviews on the platforms that matter, which is an operational change more than a budget item. Directory listings and entity consistency work is largely a one-time audit and correction effort with ongoing maintenance. Editorial coverage and third-party placements require either a PR program, a content partnership strategy, or both. The honest answer is that second-ledger work is not cheap in time, even when it is not expensive in direct spend. The reason most organizations have not done it is not cost. It is that the work is slower, less controllable, and harder to scope than a content sprint, which makes it easier to defer.

The underlying signal classes are similar across platforms: entity consistency, third-party corroboration, and prominence signals including reviews. The retrieval behavior differs. ChatGPT draws on Bing-indexed results and applies a filtering and synthesis layer that tends to favor fewer, more editorially trusted sources. Perplexity retrieves broadly in real time and cites more sources per answer. Google AI Overviews draw on Google's own index and ranking signals, which means Google Business Profile prominence and local signals carry more weight there than on the other platforms. An organization that builds strong second-ledger signals across review platforms, directories, and independent coverage will be better positioned on all three, because the underlying evidence is the same. Platform-specific optimization is less important than building the corroboration base that all three draw on.

No major AI assistant platform currently offers sponsored recommendation placement in the way paid search inserts ads into results. Organic responses on these platforms reflect the content and signals they retrieve rather than paid positions, and none of them currently documents a way to buy a recommendation. If a competitor is consistently recommended, the more likely explanation is that their second-ledger presence is stronger: more reviews, more consistent entity signals, more third-party coverage, or greater prominence in the sources these platforms draw on. Paid advertising on these platforms affects ad placements, not organic recommendation behavior. An organization that suspects a competitor is benefiting from paid placement in organic results is more likely observing the gap in their own second-ledger presence than evidence of paid influence.

A company in this position should stop treating citation metrics as evidence of recommendation progress. The two are not the same measurement, and optimizing citation rates while recommendation share stays at zero means the program is succeeding at the wrong objective. The next step is not to stop publishing. Content earns citation, and citation is necessary. The next step is to redirect a portion of the effort and budget that would have gone into additional content into second-ledger work: a review generation process, a directory and listings audit, a PR or coverage program, and an entity consistency review. The content program is not failing. It is solving the first problem. The second problem requires a different kind of work, and deferring it in favor of more content is the pattern this article exists to interrupt.


 
 
bottom of page