AI Citation Tracking: Mentions Are Not Citations

AI citation tracking records which domains an engine drew on to answer a buyer's question. Each row carries the prompt that produced the answer and the date the run happened.

A mention count tells you the engine said your name. A citation resolves to a source you can open and read. That makes it the most directly traceable AI visibility signal when you need to explain what changed.

Key Takeaways

• A citation has an address: the signal points at a source you can open and inspect. If you do not own it, you still know where the work needs to happen. A mention provides no source.
• The record is the artifact, not the rate: one row per cited domain, per prompt, per run. Six fields, kept raw.
• Page-level attribution would help less than most teams expect. You cannot commission a page on a domain you do not own. The domain is already the level at which the work happens.
• One run sets the baseline. Two runs show movement. A single read has nothing to move against, so it cannot carry a trend into a QBR.
• High visibility does not require citations to your own domain: Klarivo's 2026 benchmark put Trello second in project management on 20.4% share of voice, with trello.com cited 12 times across 1,165 answers.

A Mention and a Citation Are Different Things

A mention is your brand name appearing in the text of an answer. A citation is a source the engine names as the basis for that answer. The three-way split between mentions, citations, and descriptions belongs to the page on correcting the source rather than the answer.

That page defines all three, then goes to work on the third. This article takes the second one forward.

A citation is the only one of the three that arrives with an address attached. Everything below is about what you do with the address.

Why Only One of Them Is Traceable

Trace a mention and you run out of road inside a sentence. The answer said your name, the session closes, and the text points nowhere.

LLM citation tracking works because a citation outlives the session. The domain is still there tomorrow. So is whatever it publishes about your category.

That fourth row is the whole argument. A falling mention rate tells you the brand appeared less often. A citation list losing a domain gives you a source to investigate, and somebody can read it this afternoon.

Citation rate also sits inside a wider reporting set. What each number actually tells you is settled on its own page. This article is the record underneath one of those numbers.

Traceability stops short of proof, and saying so protects the number. The engine tells you it drew on a domain. It stays quiet about which sentence it took, and about how much weight that source carried.

What you can defend is narrow: the source was named, on that prompt, on that date. That is enough to build on. It is not enough to claim the citation caused anything.

What a Citation Record Should Contain

A citation record stores one row per cited domain, per prompt, per run. The unit is the source, not the answer. Get that wrong and an answer naming six domains collapses into a single row.

The manual audit records four fields per answer and works at answer level. It is the right first week of this work. Citation tracking goes one level down.

The sources stop being a list inside a cell. They become rows you can count, sort, and compare across months.

The Six Fields

The example values show the shape of a row. They illustrate the schema rather than reporting a run.

Six fields is a floor. Add a seventh only when you can name the decision it changes.

Why the Prompt Text Is a Field

Teams store a prompt ID and park the wording in another tab. Three months later nobody can reconstruct what PM-14 asked. The prompt is the experiment, so it belongs in the row with its result.

Wording moves the domain list more than most people expect. "Best payroll software" and "payroll software for a 40-person agency" pull different sources.

The second invites a segment answer and the first invites a league table. Storing the wording verbatim is what separates a real change from a rephrased question.

Storing the Date and the Run Conditions

A cited domain with no date is trivia. The date turns a list into a series, and the series is the only part a QBR can use.

Record the session state beside it. Three conditions matter: logged out or logged in, the country, and whether you took the first answer or a follow-up.

A run filled partly with signed-in rows is not comparable to one conducted entirely while logged out. **The same discipline explains where a tool's numbers actually come from.** It applies whether the run is yours or a vendor's.

Building a Citation Baseline

A citation baseline is the first run of a frozen prompt set. The second run creates the first comparison. Space them far enough apart that something could plausibly have changed.

One run gives you a dated baseline. The second turns it into a measure of change.

Keep this strictly about citations. The unit you record is a cited domain against a dated prompt, not a visibility score. Mixing the two produces a number nobody can decompose later.

Reusing the Prompt Set You Already Have

If you already run AI visibility reporting, the prompt set exists. Do not build a second one. Freeze it, number it, and leave it alone for a quarter.

A set that changes between runs measures your editing rather than the engines. The mechanics of building a prompt set worth tracking sit on their own page. Thirty questions you are willing to leave alone is enough to start.

What the First Run Gives You

Sort the cited domains by how many answers cited them. That ranked list is the first useful output, and it usually surprises the team that commissioned it. Three groups tend to appear:

• Directories and review sites: often among the highest-frequency rows in B2B categories.
• Category publishers: trade press and independent sites with real coverage depth.
• The long tail: domains cited once, which remain provisional until another run shows whether they return.

Count distinct answers citing a domain, not total citations. One answer naming a source three times should not outrank three answers naming it once.

Judging your own list needs a reference point. Klarivo's 2026 benchmark covers six B2B software categories on one method. It is a reasonable place to see what a good number looks like.

What the Second Run Shows

Run the same set a month later. Compare the domain lists, not the totals. A citation rate that moved four points is noise until you can name the rows behind it.

Three changes are worth writing down:

• A new domain appeared: the engine cited a source that did not surface for you last month.
• A domain disappeared: its page may have changed, the engine’s source selection may have changed, or the difference may be run-to-run variation.
• A domain remained in the citation list while your brand disappeared from the answer: one of the most actionable changes in the record.

Reading Citation Sources at Domain Level

The domain column is the most directly actionable output in citation reporting. The question underneath every citation report is how to get cited by AI at all. A ranked domain list gives you somewhere to start by naming the sources engines already cite in your category.

Reading the Domains You Own

Your own domain in the list confirms that the engine found and cited you on that prompt.

It is the easiest row to move and the least informative one. Publishing more of your own pages runs into a ceiling fast. That ceiling has its own page on why on-site work hits a ceiling.

Reading the Ones You Do Not

Klarivo's own 2026 benchmark makes the point better than an argument does. Across 35 brands in six categories, share of voice correlated +0.72 with citations to a brand's own domain. Against Domain Rating it correlated +0.47.

The correlation is strong and not deterministic, which is where the interesting rows live. Trello drew 785 mentions with 12 citations to its own domain, and SiteGround drew 540 with 27. Wrike and GoDaddy run the other way on both counts.

Trello’s 12 own-domain citations cannot account for 785 mentions on their own. Third-party sources are likely carrying much of that visibility.

That pattern is common across the benchmark. A review-site page comparing you against a competitor is a citation neither of you owns. It is often the highest-frequency domain in the whole record.

Which competitors you name changes how you read that list. That is why the denominator decides the number in any share calculation built on top of it.

The Resolution Ceiling and Why It Holds

Engines differ in what they disclose. Perplexity renders a source link for every answer and says its citations reveal exactly where an answer comes from, so page URLs are available there. Other surfaces name a source in prose, and some cite nothing at all.

Record the domain because it is the level that survives that variation. A page URL you can capture on one engine and not another is not a column you can compare across a month.

It costs less than it sounds. A community forum in your domain list tells you the engine is drawing on public discussion, and how a thread becomes citable is a question you answer at community level. The work is earning presence on the domain either way.

What Klarivo Monitor Reports

Klarivo Monitor runs this record at scale across five named engines: ChatGPT, Claude, Perplexity, Gemini, and Grok. Each is tracked independently, which matters here, because the same prompt returns different sources on different engines.

The reporting surface is Top 10 Citation Sources and it lists domains. It updates on a schedule you control. It exports and filters alongside the mention and share-of-voice figures from the same runs.

Monitor reports at domain level for the same reason the record above does, and that is deliberate. A page-level number that only some engines can produce is not a number you can trend across all five. The schema in this article and the surface in the product share a shape because they answer to the same constraint.

The dashboard stops at the list. Klarivo’s managed service handles what comes next: building a presence on the domains you do not own.

If you have a citation list and no plan for the domains on it, book a 15-minute Klarivo discovery call at a time that works for you. Your booking is confirmed immediately.

Frequently Asked Questions About AI Citation Tracking

Is an unlinked brand mention inside a cited page a citation?

No. The page is still a cited source, but the brand mention inside it is not an AI citation. Record the cited domain, then set Brand named to No if the engine’s answer did not name you.

The two rows lead to different work. A source cited in an answer that does not name you is a domain to investigate. A source supporting a wrong description is a page to correct.

What should I do when two engines cite completely different domains for the same prompt?

Treat it as the expected result rather than a fault in the record. Keep both rows, then read the overlap separately from the union.

Domains appearing across three or more engines are cross-engine targets worth investigating. Domains appearing on one engine tell you about that engine's retrieval. Watch those before you act on them.

Do backlinks still matter if AI citations are the signal I am tracking?

Yes, and they do different jobs. A link can pass authority to your domain. A citation puts a source in front of a buyer inside an answer.

The two lists overlap less than most SEO teams expect. Klarivo's 2026 benchmark found share of voice tracked own-domain citations more closely than Domain Rating. That is a reason to keep both.

Can I reconstruct a citation baseline for a period I did not track?

No. Engines provide no complete queryable record of past answers for you to pull. Chat history helps only if you ran the prompts and kept the sessions.

Unless you did, re-running old prompts returns today's results. Start the record now, and accept that your first defensible comparison is a month away.

Should citations from Google's AI Overviews go in the same record as ChatGPT citations?

Yes, in the same table, with the engine field doing the separating. One schema is what lets you compare surfaces at all.

The engine column exists so you can filter down when a single surface behaves oddly. Splitting the record by surface is the fastest route to four spreadsheets nobody joins.