Mentions Get You Named. Citations Get You Credited.

An AI mention is when an answer engine names your brand. An AI citation is when it credits a source — your URL, a footnote, a linked title, or an explicit attribution. You can be mentioned without being cited. You can be cited without a flashy brand shout-out. Treating them as the same KPI is the #1 reporting failure in AI visibility programs.

This spoke sits under the Answer Engine Optimization playbook. Pair it with What Is AI Visibility for the outcome definition. The logging method lives in Measuring AI Search Visibility.

The short answer

• Mention = named in the answer text. Citation = credited as a source the user can open.
• Mentions help awareness. Citations help trust and click paths.
• High mentions + low citations usually means soft brand presence without extractable proof.
• Dashboard both columns separately. Never merge them into one “visibility %.”
• Fix citeability when mentions outrun citations. Fix brand-in-the-lead when citations outrun names.

What’s the difference between an AI mention and an AI citation?

A mention is a string. A citation is a credit. Schema.org defines citation as a reference to another creative work — a publication, web page, or article — not a name-drop in prose. Answer engines borrowed that split even when their UI hides it.

	Mention	Citation
Signal	Brand string appears in the answer	Domain or document used as a source
Buyer experience	“I’ve heard of them”	“Here’s proof / a place to go”
Typical UI	Plain text name	Link, footnote, source card
Logging column	mentioned: yes/no	cited_url: …
Over-report risk	High if you screenshot once	Lower if you require a URL

Operator test: if you delete every link and source chip from the answer, what remains naming you is a mention. What disappeared was the citation surface.

Semrush’s own toolkit treats the two as different metrics. Mentions count prompts where a brand is included in the response. Citations count responses that cite your domain as a source. If your vendor dashboard already splits them, do not re-merge the columns in the slide you send to leadership.

The four states every answer can land in

Every panel run lands in one of four cells. If your sheet only has a single “visible?” checkbox, you are collapsing a 2x2 into a lie.

State	Mention	Citation	What it actually means
Named and credited	Yes	Yes	Full win. Brand in the sentence; URL in the source list.
Mention only	Yes	No	Awareness without a landing path. Common on recommendation lists.
Citation only (ghost)	No	Yes	Credit without brand memory. Common on how-to and definition answers.
Absent	No	No	Miss. Competitor or publisher may own both cells.

• Do not call mention-only a “content win.”
• Do not call citation-only a “brand win.”
• Do not call a miss a “tie” because a competitor was also absent.
• Write a one-line note when the answer is messy. Silent upgrades are how vanity reports get born.

The rarest cell in published research is the full win. In the Semrush / Kevin Indig ghost-citation study (3,981 domain appearances, 115 prompts, 14 countries, four engines, June 2026), only 13.2% of appearances were both cited and mentioned. That is the cell revenue teams think they are buying when they say “AI visibility.”

Ghost citations: cited, never named

Kevin Indig coined ghost citation for the gap: the model uses your page as a source link and never says your name in the answer. Readers can click through. Most will not. They leave with the method, not the brand.

The same study split every appearance into three buckets:

Bucket	Share of appearances	What the user saw
Ghost citation	61.7%	Source link, no brand name in the answer
Mention without citation	25.1%	Brand named, no owned URL credited
Cited and mentioned	13.2%	Name plus source link

Citation rate in that dataset was nearly double mention rate: 74.9% of appearances included a citation; only 38.3% included a brand mention. If you only track citations, you will over-read authority. If you only track mentions, you will over-read fame.

That is not a vibe. It is why this spoke exists as its own page instead of a footnote in the measurement post.

Can I be mentioned without being cited?

Yes — constantly. About a quarter of appearances in the Semrush study were mention-only. You will see it on live panels every week.

1. ChatGPT recommends “Acme HVAC” in prose with no sources shown.
2. An Overview names a category leader verbally while citing a roundup that never links you.
3. A comparison answer lists your brand among “options” while sourcing a competitor’s blog.
4. Gemini names you from model memory and never attaches a URL. In the same study, Gemini mentioned brands in 83.7% of appearances and cited a source only 21.4% of the time.

Mention-only is not worthless. It is awareness. Models learn you belong in a category when your name keeps showing up next to the right nouns. It does not give the buyer a page to evaluate, a proof point to screenshot, or a path into your funnel.

Log it as mentioned: Y, cited_url: —. Then ask which third-party page got the footnote instead of you. That URL is the real competitor for the next ticket — not the brand that was also named.

Can I be cited without a brand mention?

Yes. That is the ghost-citation cell, and it is the majority cell in the Semrush dataset. ChatGPT in that study cited a domain 87% of the time it appeared and mentioned the brand only 20.7% of the time. The answer looks like a footnoted paper. Your URL is in the appendix. Your name is not in the lead.

Also possible on Perplexity: the product cites your how-to URL, paraphrases your method, and never says the company name in the first sentence. Perplexity’s own help hub describes the product as delivering answers “with sources and citations included.” Sources can be present while the brand string is absent. Log that as a citation win plus a brand-mention miss.

Publisher and aggregator domains get this pattern the most. In the same study, medium.com was cited 16 times and never named. Wikipedia, academic hosts, and review roundups get used as raw reference. Consumer brands with strong public identities get the reverse: named more often than linked.

If you sell a service, a ghost citation is a half-win. Someone could click. Most people will remember the steps, not the studio.

Which matters more for pipeline — mentions or citations?

Depends on the job of the prompt. Do not pick a favorite KPI and apply it to every bucket.

Prompt type	Prefer	Why
Category recommendation	Mentions + citations	Name gets you shortlisted; cite proves you
How-to / problem	Citations	Buyer needs a trustworthy method URL
Comparison	Both	Name for consideration; cite for criteria pages
Brand / reputation	Mentions (accurate)	Wrong facts here are worse than silence
Local service	Mentions (NAP-consistent) + Maps	Click path often leaves the chat

Rule of thumb for revenue teams: citations move evaluation; mentions move awareness. If your funnel is cold outbound and brand-new, chase accurate mentions first. If inbound already knows the category, chase citations on the pages that close.

The Semrush study backs the split by intent. Informational queries (“what is,” “explain,” “how does”) posted an 89.3% citation rate and an 18% mention rate. Comparative queries (“best,” “vs,” “recommend”) produced a 43.3% mention rate — 2.4x more brand mentions than informational. How-to sat at 42.8% mention rate. Commercial queries (pricing, buying) sat at 35.6% mention rate and 84.4% citation rate.

If you average those buckets into one “AI visibility” number, you will celebrate how-to ghosts and miss the recommendation prompts that actually shortlist vendors.

How ChatGPT, Perplexity, and AI Overviews show the two signals

Interfaces change. The logging habit should not. Record product-native evidence: screenshot, cited URLs, and a boolean for brand named. Do not invent a universal “citation score” across products.

Product	Mention pattern	Citation pattern
ChatGPT	Brand in prose; sometimes no sources	Inline citations when search is on; otherwise a Sources panel
Perplexity	Brand in answer body	Numbered footnotes plus a source list on almost every answer
Google AI Overviews	Brand in summary text	Supporting link modules; eligibility tied to indexing and snippets
Gemini (in the Semrush set)	High name rate, low link rate	Often talks from memory; citation is the exception

OpenAI’s ChatGPT Search help is explicit: responses that use search may include inline citations; if they do not, you click Sources beneath the response to open cited links. That “may” is the whole measurement problem. A ChatGPT answer that names you with no Sources panel is a mention. A ChatGPT answer that lists your URL in Sources and never says your name is a citation. Same product. Two columns.

Google’s documentation is equally blunt about links, not names. AI features and your website says AI Overviews and AI Mode “surface relevant links” and that eligibility to appear as a supporting link requires the page to be indexed and snippet-eligible. Google’s May 2025 Search Central note adds that AI Overviews “display links in a range of ways.” Range of ways means your logging rubric has to accept source chips, numbered modules, and carousel cards — not one screenshot shape.

Semrush found almost no overlap between the brands ChatGPT cited and the brands Gemini named for the same prompts. Treat the products as different scoreboards.

Sources consulted vs citations shown

There is a third layer underneath the two you report: the model can retrieve a URL, use it, and never show it.

OpenAI’s web search tool docs separate the fields on purpose. Inline citations are “the most relevant references.” The sources field returns the complete list of URLs the model consulted. “The number of sources is often greater than the number of citations.”

Google describes a similar gap without giving you the hidden list. AI Overviews and AI Mode may use query fan-out — multiple related searches across subtopics — then show a smaller set of supporting links than the system retrieved.

Layer	Visible to the buyer?	What you can log	What you cannot honestly claim
Consulted / retrieved	Usually no	Rarely, unless you have API sources	“They used our page” from a screenshot
Cited	Yes	URL, title, position in the source list	“They named us”
Mentioned	Yes	Brand string or clear alias	“They credited our URL”

For a marketing panel, log what the buyer saw. Consulted-but-uncited is real. It is not a KPI you can defend in a Monday meeting unless you have the API payload. Do not upgrade a hunch into a win.

What content wins citations vs what wins mentions?

Citations favor extractable owned pages. Mentions favor third-party corroboration and a name the model already trusts. You need both layers. They are not the same content calendar.

Goal	Content that tends to win	Why models grab it
Citations	Definitions, tables, numbered methods, original stats	Extractable, attributable passages
Mentions	Roundups, Reddit threads, reviews, “best of” lists	Social proof and entity co-occurrence
Both	Criteria pages + case studies with named outcomes	Brand + quotable structure

Semrush’s share-of-voice explainer draws the same three-term split operators keep collapsing:

Term	What it measures	Example
AI mention	Brand appears inside the generated answer	ChatGPT names you in a list of tools
Citation	Linked reference to your content	The answer links your pricing page for a claim
Source	Page the system retrieved to build the answer	A review the model pulled from
AI share of voice	Your mentions relative to named competitors	18% of category mentions, not 18% of the internet

Mentions without owned proof pages create a hollow brand — famous in chat, nowhere to land. Citations without a name create a useful ghost. Fund the layer that is actually empty.

Query phrasing changes the ratio

Topic is not enough. The same topic asked two ways can flip the mention/citation mix.

The Semrush study measured prompt length from 16 to 98 characters (about 60 on average). Short, conversational queries produced brand mention rates near 100%. Long, structured prompts produced mention rates of 2%–3% — a 30x–50x gap — while triggering more citations. “Should I lease or buy a car for my business?” named brands. The same question padded with extra context and framing produced ghosts.

Prompt shape	Typical mention rate in that study	Typical citation behavior	How to use it on a panel
Short, conversational	Near 100%	Fewer footnotes	Brand / recommendation coverage
Long, structured	2%–3%	More source links	Citeability and extractability
Comparative (“best,” “vs”)	43.3%	Mixed	Shortlist + proof
Informational (“what is”)	18%	89.3% cited	Ghost-citation watchlist

• Freeze both a short and a long variant for each money topic.
• Do not retire the long variant because “it never names us.” That is the citeability test.
• Do not celebrate the short variant as a content ROI win. It is a name test.
• Re-phrase quarterly from sales notes, not from whatever ChatGPT suggested this morning.

If your panel is 40 long how-tos, you will look well-cited and unknown. If it is 40 “who should I hire?” one-liners, you will look famous and unlinked. Build both.

Intent buckets do not share a scoreboard

Averages hide the job. Split the panel the way buyers actually ask.

Bucket	What “good” looks like	False comfort
Category / recommendation	Named in the shortlist; cited on the criteria page	Named in a joke list with a competitor’s URL
Comparison	Named next to the right peers; your comparison URL cited	Named, but the cited table is a rival’s
How-to / problem	Your method URL in Sources	Your method paraphrased, publisher cited
Local	NAP-consistent name; Maps / GBP path	City name only, wrong phone
Brand / reputation	Accurate facts, even with no link	Mentioned with a hallucinated price or office

Weight or separately report the buckets. A high citation rate on vanity how-tos with zero recommendation inclusion is a false comfort. A high mention rate on brand queries with rotting facts is a liability.

Country is a bucket too. In the Semrush set, India and Sweden saw brand names in 50% of answers. Italy, Brazil, and the Netherlands sat at 18%–22% mention rate even while citation rates ran 82%–94%. If you operate in more than one market, do not apply a single “we are visible” claim globally.

How do I log messy answers without upgrading them to wins?

Answers are messy. Use a consistent rubric and refuse silent upgrades.

1. Brand string exact or clear alias → mention = yes.
2. Your domain in sources / footnotes / link modules → citation = yes.
3. Competitor domain only, you named → mention without citation.
4. Your method paraphrased, no brand, your URL cited → citation without mention.
5. Hallucinated facts about you → accuracy fail regardless of mention or cite.
6. Alias only (“the Nashville studio,” no legal name) → mention = no unless the alias is a registered brand you already mapped.
7. Homepage cited for a claim that lives on a spoke → citation = yes, note the wrong URL.
8. Source chip present, URL blocked or 404 → citation = yes, accuracy / hygiene fail.

Edge cases get a note, not a promotion to “win.”

Messy pattern	Mention	Citation	Note to write
Name in a list, Sources empty	Y	N	mention-only
Your URL in Sources, name absent	N	Y	ghost citation
Competitor URL, your name	Y	N	gap URL = competitor
Right method, wrong year / price	Y or N	Y or N	accuracy fail
Trademark misspelling	N	as observed	do not “fix” the log

If two operators would score the same screenshot differently, the rubric is still too soft. Write the rule down. Re-score last week’s panel against it before you change the site.

Should my KPI dashboard track both?

Yes. Semrush’s AI visibility reporting guide says the same thing in vendor language: when AI Overviews and AI Mode mix into regular search data, report mentions, citations, and cited pages separately from organic traffic. Do not wait for a tool to invent a blended score you will later have to unwind.

Minimum columns per panel run:

• Date
• Product
• Prompt ID
• Prompt bucket (recommendation / how-to / comparison / local / brand)
• Brand mentioned (Y/N)
• Cited URL(s)
• Citation position if the UI orders sources
• Competitor brands named
• Competitor URLs cited
• Fact accuracy flag
• Screenshot / archive link

Weekly rollup: mention rate, citation rate, share of voice, top gap URLs. Monthly: first-source rate and AI referral sessions. Full instrumentation lives in Measuring AI Search Visibility.

Mention rate without citation rate hides hollow visibility. Citation rate without mention rate hides brand-blind how-to wins. Accuracy without either hides a quiet disaster: you are invisible and still being misquoted on the prompts you do win.

Failure mode: the mention vanity report

What breaks: marketing reports “87% AI visibility” because the brand string appeared in 87% of ChatGPT answers — all unlinked, many inaccurate, zero owned URLs cited.

What it costs: budget shifts to PR name-drops while the site remains unquotable. Sales still hears competitors when buyers ask “who should we hire?” The board thinks the AI problem is solved. The pipeline does not.

What you do instead:

• Split columns: mention rate / citation rate / accuracy.
• Require a cited URL before calling a run a “win” for content ROI.
• Attach every content ticket to a prompt ID and a target citation URL.
• Report recommendation-bucket mention rate separately from how-to citation rate.
• Kill any slide that averages ChatGPT, Perplexity, and AI Overviews into one percentage.

Bravery is not a dashboard that only counts name strings.

The inverse failure is quieter and just as expensive: a content team celebrates “citation rate up 40%” on definition pages while recommendation prompts still name two competitors and cite a G2 roundup. That is a ghost-citation program, not a demand program.

How do I fix high mentions but low citations?

You are famous in the chat and missing from the footnote. Run this ladder in order. Do not start at “write more blogs.”

1. Inventory — 20 recommendation prompts; count mention-only vs cited.
2. Gap URLs — list every URL that gets cited instead of you.
3. Quote test — can a stranger lift 40–80 words from your page as a standalone answer?
4. Truth layer — About, schema, and directory facts match.
5. Displace — publish or earn the page type the model already prefers (comparison table, checklist, stats).
6. Re-test — same prompts, same products, next week.

Skip volume until steps 1–3 are done. Volume without extractability produces more mentions of other people.

Check	Pass	Fail
Lead sentence states the answer	Yes, in words a buyer would repeat	Opens with company history
Table or numbered method on the page	Present and unique	Prose wall
Brand name in the title or H1	Present	Generic “ultimate guide”
Claim has a source the model can re-cite	Original number or named method	Recycled listicle

Mentions outrunning citations is usually an on-site extractability problem, not a PR shortage. You already got the name. You have not given the model a passage it is willing to credit.

How do I fix high citations but low mentions?

You are the footnote and not the noun. That is the ghost-citation pattern. Different ladder.

1. Confirm the cell — citation = yes, mention = no, on the same prompt, same product.
2. Open the cited URL — is your brand in the first 150 words, or only in the footer?
3. Name the entity in the lead — “Spurlock Studios defines X as…” beats “Teams often define X as…”
4. Add a comparison or criteria block — comparative prompts mentioned brands 2.4x more often than informational ones in the Semrush set.
5. Check aliases — legal name, product name, and the name sales uses must match the page.
6. Re-test the short conversational variant — long how-tos will keep ghosting you even after the lead is fixed.

Check	Pass	Fail
Brand or product in the opening answer	Named once, naturally	Method only, brand in the byline
Page type matches a naming prompt	Comparison, alternative, “best for”	Pure glossary
Off-site pages also use the same name	Directories and profiles agree	Three spellings in the wild
Cited URL is a money page	Service / criteria / research	Orphan blog with no entity

Do not hire PR to “get mentioned” if the page the model already cites refuses to say your name. You will buy more ghosts.

Decision table: what to fund next

Use the table in planning meetings so creative and SEO stop arguing from different scoreboards.

Pattern in the log	Fund first	Do not fund yet
Low mentions, low citations	Entity consistency + corroboration (PR / directories)	Large blog volume
High mentions, low citations	Answer-first rewrites + tables on money pages	More unlinked brand seeding
High citations, low mentions	Branding in titles and leads; owned criteria pages	Random guest posts
Mentions with accuracy fails	Fact sheet + cleanup before growth	Aggressive PR push
Citations only on vanity how-tos	Shift panel weight to recommendation prompts	Celebrating “content wins”
ChatGPT cites, Gemini names, Overviews ignore	Per-product tickets, not a blended campaign	One “AI SEO” retainer with no product split

If two teams cannot point at the same row, you do not have a strategy disagreement. You have a logging disagreement. Fix the sheet first.

Sample log rows (copy into your sheet)

Three rows teach a junior marketer more than a 40-slide “AI visibility” deck.

date	product	prompt_id	bucket	mentioned	cited_url	note
2026-01-12	perplexity	rec-03	recommendation	Y	competitor.com/best-x	mention-only; gap URL
2026-01-12	chatgpt	how-07	how-to	N	yours.com/guide	ghost citation
2026-01-12	aio	cmp-02	comparison	Y	—	mention only
2026-01-12	gemini	rec-03	recommendation	Y	—	named, no source
2026-01-12	chatgpt	rec-03	recommendation	Y	yours.com/criteria	full win

Notice rec-03 is not one result. It is four product rows. Averaging them into “75% visible” is how the vanity report starts.

• One row per product per prompt per date.
• Prompt ID is stable. Wording can drift; the ID cannot.
• cited_url is a real URL or an em dash. Never “yes.”
• Notes use the four-state language: mention-only, ghost, full win, miss, accuracy fail.

Weekly ritual (15–30 minutes once the panel exists)

1. Run the frozen panel on priority products.
2. Fill the two booleans + cited URLs.
3. Flag any accuracy fail for cleanup before you chase new inclusion.
4. Add new gap URLs to the leaderboard.
5. Open at most three tickets: one extractability, one corroboration, one accuracy.

If the ritual takes three hours, your panel is too big or your logging is too theatrical.

Minute	Action	Output
0–10	Re-run the frozen prompts	Fresh screenshots
10–20	Score mention / citation / accuracy	Updated sheet
20–25	Diff vs last week	New gap URLs, lost cites
25–30	File ≤3 tickets	Owners and due dates

Do not add prompts during the ritual. Prompt-list edits are a monthly job. Mid-week additions are how teams “improve visibility” by measuring easier questions.

How this distinction feeds a visibility audit

In a Spurlock Studios visibility audit we separate mention and citation from day one, then map which layer is broken — entity, content extractability, or off-site corroboration. AI visibility is the scoreboard. This spoke is the unit definition. The AEO playbook is the system those units plug into.

If you cannot say, for a given prompt and product, whether you were named, credited, both, or neither, you are not measuring AI visibility. You are collecting screenshots.

I have spent 20,000+ hours on agentic systems and have been SEO-certified since 2021. The teams that get this distinction ship fewer posts and make better tickets. The teams that skip it buy a dashboard that applauds name strings.

FAQ

Do unlinked brand mentions still help?

Yes for awareness and entity co-occurrence — models learn you belong in a category, but they help less for evaluation and click-through than a real citation. Count them; do not celebrate them as content ROI. In the Semrush / Indig dataset, mention-only was 25.1% of appearances; treat that slice as a name test, not a page test.

Do citations always include my brand name?

No. A product can cite your URL while paraphrasing the method without saying the company name in the lead. Semrush found 61.7% of citations were ghosts. Log citation and mention as separate fields so you catch that pattern instead of calling every footnote a brand win.

How do ChatGPT and Perplexity show citations differently?

Perplexity almost always exposes numbered sources plus a source list. ChatGPT may answer in prose, with inline citations only when search is on, and a Sources panel when inline marks are missing. Google AI Overviews mix summary text with supporting link modules. Log product-native evidence instead of forcing one UI model.

Should my KPI dashboard track both?

Yes. Mention rate without citation rate hides hollow visibility. Citation rate without mention rate hides brand-blind how-to wins. Track both, plus accuracy, and keep recommendation prompts on a separate rollup from informational ghosts.

What content wins citations vs what wins mentions?

Citations favor extractable owned pages: definitions, tables, steps, original numbers. Mentions favor third-party corroboration and a name the model already trusts: roundups, reviews, community threads, comparative “best” prompts. You need both layers. They are not interchangeable calendar items.

How do mentions relate to digital PR?

PR increases the chance you are named in the sources engines retrieve. Citations still need a primary page worth crediting. Buy PR after (or with) citeable assets — not instead of them. If the model already cites you and still will not say your name, fix the lead on the cited URL before you buy another listicle.

CTA

Stop reporting “visibility” as a single percentage. Split the columns, then fix the gap.

Lane: /visibility · Next step: visibility audit