There is a comforting version of AI visibility where the job is simply to get your brand mentioned.

Ask a model about your category. See your name in the answer. Celebrate.

The problem is that a mention can mean several different things. It can mean the model knows you exist. It can mean you are a weak alternative. It can mean you are included because everyone in the category gets name-dropped. It can mean the answer is outdated. It can mean the model repeats your name while citing somebody else as the source.

A citation is different. A citation means the answer is willing to point at a source. A recommendation is different again. A recommendation means the answer is steering the user toward a choice.

If your measurement collapses all of that into “we appeared,” you are flying with the most important instruments switched off.

The four visibility events.

AI visibility should be split into at least four events.

Four events, four questions

FIG. 01

A mention Awareness: The answer names your brand, product, location, or page.
A citation Evidence: The answer links to or otherwise attributes information to a source associated with you or relevant to you.
A recommendation Commercial: The answer presents you as a suitable option for the user's need.
A top pick Decisive: The answer makes you the leading choice, not just one item in a list.

These events can overlap. They are not interchangeable.

Those events can overlap, but they are not interchangeable.

You can be mentioned without being cited. You can be cited without being recommended. You can be recommended through someone else’s article rather than your own site. You can be the top pick in one model and invisible in another.

That is why AEO reporting should not stop at “brand appeared: yes/no.” The useful question is what kind of appearance happened and what it implies.

A mention is awareness. A citation is evidence.

A mention tells you the model has some representation of your brand. That matters. If a model never names you for relevant category prompts, you have an awareness or retrieval problem.

But a mention alone does not tell you whether the answer trusts your content. It may be relying on general training memory, an old listicle, a marketplace profile, a directory, a review site, or a competitor’s comparison page. The model can know your name and still not treat your site as a source.

A citation is a stronger signal because it gives you a path. Be precise about what it proves: a visible citation is evidence that the product chose to show that source under its own citation rules. It is not proof that the source was the only one retrieved, or the most influential one. That source can be your own page, a documentation page, a review platform, a partner page, a news article, a Reddit thread, a YouTube transcript, or a competitor-owned comparison.

That source path matters because it tells you where authority is sitting.

If Perplexity cites your documentation when explaining your product, your owned content is doing work. If it cites a third-party roundup, you have earned media doing work. If it cites a competitor comparison, you may be present in the answer but not in control of the frame. If it cites outdated directory data, your facts are at risk.

Mentions tell you whether you are in the model’s world. Citations tell you what the model trusts enough to show.

A recommendation is the commercial event.

For many teams, the business value is not the mention or the citation. It is the recommendation.

If a buyer asks “which tools should I evaluate?” or “where should I eat tonight?” or “what are the best hiking boots under $150?”, the answer is not merely informational. It shapes a shortlist.

That is the commercial event AEO is trying to measure. Did the model recommend you? Did it recommend a competitor? Did it recommend the wrong product line? Did it hedge? Did it call you expensive, limited, niche, enterprise-only, outdated, or hard to use? Did it include a caveat that would make a buyer pause?

This is where sentiment and accuracy matter. A model can mention you in a negative frame. It can cite you while recommending someone else. It can recommend you for the wrong buyer, creating poor-fit demand.

Good AI visibility reporting treats the answer like a sales conversation that happened without you in the room. The question is not just “were we there?” It is “what did the buyer hear?”

Why citations are becoming a measurement layer.

The web is moving toward more zero-click behavior. Pew Research Center analyzed March 2025 browsing data and found that Google users who encountered an AI summary clicked a traditional search result in 8% of visits, compared with 15% of visits without an AI summary. SparkToro’s zero-click research has also tracked the broader decline in searches that send clicks to the open web.

You can debate the exact magnitude by dataset and methodology. The strategic direction is clear: more answers are happening before the click.

That makes citations more important, not less. If fewer users click through, the source shown inside the answer becomes a visible trust signal. It is one of the few places where a brand can still see which pages influenced the interaction.

Adobe’s data shows the click-based side is growing too. Generative AI referrals to retail sites grew sharply from 2024 into 2025 and continued to grow in 2026. But referral visits are the part that reaches your analytics. Citations reveal the part that happens inside the answer.

The most serious teams will measure both.

Cited versus retrieved.

There is another distinction that matters: a source can be retrieved without being cited.

In a retrieval-based system, the model or search layer may pull a set of candidate documents, use some of them to construct an answer, and then show only a subset as visible citations. That means a page can influence an answer without receiving a visible citation. It also means a visible citation is not the complete map of what the system considered.

This distinction is not always exposed by AI products, but it is important conceptually. If a model retrieves your page often but rarely cites it, the page may be useful but not citation-worthy. If it retrieves competitor pages and cites them, the competitor has stronger source fit. If it retrieves third-party articles and cites those, the answer layer may trust external validation more than brand-owned claims.

Omni Impact’s Sources view is built around this exact problem: which sources are being used, which are being shown, and how often your brand is connected to them.

What changes a citation rate?

There is no magic checklist, but the patterns are becoming clearer.

The generative engine optimization research paper by Aggarwal and colleagues introduced a benchmark of 10,000 queries and found that optimization methods could improve source visibility by up to 40% across diverse queries. The paper found that the value of different tactics varied by domain, which is the part marketers should remember. There is no universal trick. There are source-quality patterns.

The practical patterns are:

Put answerable facts in crawlable text.
Add credible citations when making factual claims.
Include statistics where they genuinely support the point.
Use clear structure so passages can be extracted.
Create pages that answer specific buyer questions.
Keep facts consistent across owned and third-party sources.
Earn mentions from sources models already trust in your category.
Avoid keyword stuffing and thin “AI bait” content.

The goal is not to manipulate an answer engine with decorative citations. The goal is to make your page a better source than the alternatives.

How to audit mentions and citations manually.

You can start without software. It will not scale, but it will teach you what to look for.

Pick 20 prompts that represent real buying moments. Include a mix of:

Category prompts: “best [category] for [use case]”
Comparison prompts: “[your brand] vs [competitor]”
Alternative prompts: “alternatives to [competitor]”
Local prompts: “best [service] near [location]”
Trust prompts: “is [brand] reliable?”
Feature prompts: “which [category] tools support [integration]?”

Run them across multiple models. Save the full answer, not just the excerpt that mentions you.

For each answer, record:

Was your brand mentioned?
Was your own site cited?
Was any third-party page about you cited?
Was a competitor cited?
Were you recommended?
Were you the top pick?
Was the description accurate?
Which source seemed to shape the frame?

Then look for patterns:

You are absent from high-intent prompts.
You are mentioned but not cited.
You are cited for old facts.
Competitors are cited through listicles.
Third-party review sites define your category.
Your own pages are too vague to use as evidence.
One model likes you; another does not.

That pattern is the work queue.

How to turn the audit into action.

If you are not mentioned, build category presence. Create pages that match the way buyers ask. Publish comparison and alternative content where appropriate. Make your positioning explicit. Earn credible third-party mentions.

If you are mentioned but not cited, improve source quality. Add factual density, clear headings, supporting data, and structured information. Make sure crawlers can access the page. Reduce unsupported claims.

If you are cited but not recommended, improve fit and differentiation. The model may know what you do but not why you are a good choice. Add use cases, buyer profiles, integrations, limits, pricing context, and decision criteria.

If competitors are cited more often, inspect their source paths. Are models citing their docs, review pages, comparison pages, customer stories, marketplace listings, or press? Do not copy their content. Understand what evidence is missing from your public footprint.

If the model is wrong, fix the public facts. Update your site, profiles, docs, schema, listings, and pages that third parties cite. In AI answers, stale facts can live longer than you expect.

What to report internally.

Executives do not need a spreadsheet of raw prompt outputs. They need the shape of visibility.

A useful report includes:

Overall mention rate across core prompts.
Citation rate for owned sources.
Citation rate for third-party sources.
Recommendation rate.
Top-pick rate.
Competitor share of voice.
Accuracy issues.
Source leaderboard.
Highest-impact fixes shipped.
Movement over time.

This keeps AEO out of the realm of vibes. It becomes an operating cadence: measure the answer layer, fix the source layer, and watch whether visibility changes.

That is how Omni Impact frames the problem. Mentions, citations, sources, prompt performance, and share-of-voice are separate because they answer separate questions.

The bottom line.

A mention is not a win by itself. It is a signal.

A citation is stronger, because it shows one source the answer was willing to stand behind — though not necessarily the only one it used. A recommendation is a different event again, because it shapes the buyer’s shortlist whether or not your page was ever cited. A top pick is the event every brand wants, because the answer did not just include you. It chose you.

Keep them apart in the reporting. A team that treats a citation as a recommendation will celebrate being quoted while a competitor gets bought.

The brands that understand this distinction will not be satisfied with “ChatGPT named us once.” They will ask the better questions:

Who is the model citing? What does it trust? Are we recommended? Are we accurate? Are competitors owning the source paths? Which fix would change the answer?

That is the difference between watching AI visibility and managing it.

Mentions Are Not Citations: The AI Visibility Metric Most Teams Miss