AI Citations

AI Citations shows where AI assistants cite your site when they answer people’s questions — which of your pages they point to, how visible you are among the sources those engines cite, and what that visibility is worth in real visits and conversions. You will find it in the left sidebar, under your site, as an AI Citations group that opens into two pages:

• Analytics (/citations) — the read-only picture: your verdict, share of voice, cited sources, the per-prompt outcomes table, most-cited pages, and a site-wide Audience breakdown.
• Tracked prompts (/citations/prompts) — where you manage things: add and pause prompts, set cadence and engines, and tag the sources that actually compete with you.

It answers a question the rest of your dashboard can’t: when someone asks ChatGPT or Perplexity a question your site could answer, does the assistant cite you — and does that turn into anyone actually visiting? You tell us the questions (“prompts”) you care about, we watch the answer engines for them, and we join what we find to your on-site traffic.

The whole feature is built to be honest about how sure it is. Every figure is shown at visit grain, carries a confidence band, and is labelled directional unless it is well-sampled enough to call a measurement. We never dress a number up as more precise than it is, and we never call a co-cited domain a “competitor” until you have tagged it as one. The Methodology section at the end explains the rules in full.

[Screenshot: The AI Citations group in the sidebar, expanded to Analytics and Tracked prompts]

Reading the period

Like the rest of the dashboard, the Analytics page follows a period selector at the top. It defaults to the last 30 days, with 7-day, 30-day, and 90-day presets. Everything below — the verdict, share of voice, the outcomes table, most-cited pages, Audience — is scoped to the period you pick. If a panel looks empty, widening the period is often the fix. (The Tracked prompts page has no period control — it is a management surface, not a report.)

The Analytics page

The verdict (the headline hero)

At the top sits a single plain-language answer to “how am I doing?”. When answer engines are citing you, it reads something like:

You hold 38% of all citations on your tracked questions

with the subline Answer-engine citations and the on-site visits they drove. Just under it you will see a row of small honesty chips (the methodology label, the sample count, the confidence band) — see Methodology for what those mean.

The verdict changes to match your real situation, so it is never misleading:

• Other sources are cited for your tracked questions — you aren’t yet. Engines are answering your questions, but pointing at other sites instead of you. Open a prompt’s answer view to see who was cited in your place.
• Collecting your first results. You have added prompts, but no engine has been polled for them yet. First results land within about a day.
• No AI citations yet. You haven’t added any prompts to track. Add one on the Tracked prompts page to get started.

Note the wording: when you are not cited, the verdict says “other sources”, not “competitors”. A domain cited beside you is just a source until you deliberately tag it as a competitor — see Tagging competitors.

The stat cards

When you are being cited, three cards sit under the verdict:

• Citations observed — how many times engines cited you across your tracked prompts in the period.
• Visits to cited pages — how many real on-site visits those cited pages drove.
• Estimated contribution — the revenue we can attribute to those visits, shown only when your site has a goal that tracks revenue. With no revenue goal, this card shows Conversions instead — a plain count of goal completions.

Share of voice

The Share of voice panel shows your slice of the citation “voice” among the sources answer engines cite beside you. Your own row is highlighted and labelled (you); each row shows a count, a share percentage, a bar, and an aggregate sentiment badge (positive / neutral / negative / mixed / unscored) for how engines tend to frame that source.

There are two ways to count, toggled with the By volume / By presence buttons. They answer slightly different questions:

• By volume (the default) counts every citation. If an engine cites a source three times in one answer, that’s three. It is the best proxy for attention and click surface — but it is easier to game, because a source with a lot of content can rack up citations.
• By presence counts each source once per answer, no matter how many times it is cited there. It is the “reach” view — how many distinct answers mentioned each source — and it is harder to game.

Switching between them never reloads anything; both numbers come from the same data.

The panel also has a lens toggle: All sources (the default) and Competitors.

• All sources shows every domain cited beside you, whether you have tagged it or not. This is the honest default — most of these are just other sites the engine happened to cite.
• Competitors narrows the panel to the sources you have tagged as competitors (see Tagging competitors). If you haven’t tagged any yet, a hint points you to the Tracked prompts page. The word “competitor” only appears once you’ve made that call — we never guess it for you.

If there are more sources than fit, the panel shows the top few and folds the rest into a ”+ N others” row, so a source is never silently dropped. A note tells you how many were collapsed.

The tracked-prompts outcomes table

The outcomes table is the primary object on the page — one row per tracked prompt, ordered opportunity-first (the prompts most worth acting on rise to the top). Each row shows, per engine, whether you were cited, mentioned, or absent in the grounded answers, plus:

A Gaps filter narrows the table to prompts where your brand was absent from every grounded answer, banded by how much retrieval evidence there is — the clearest to-do list. Choose View analysis on any row to open its answer view.

Early on, when little AI-referral traffic has reached your site, a note explains that citations are your leading indicator and visits typically follow — the ranking leans on prompt intent and citation strength, not traffic, until the traffic arrives.

Most-cited pages

Below the table, Most-cited pages joins the pages engines cite to what those pages actually did on your site. Each row is one of your pages:

The confidence band

Matching a real visit to a specific citation is probabilistic — we line up AI-referred visits with the page an engine cited, and the strength of that match varies. So each row carries a confidence band — low, medium, or high — rather than a bare exact number. A low band is shown plainly and never hidden or inflated to look certain. See Methodology for the bands in detail.

Dashes, never fake zeros

Where there is no goal data to compute a conversion rate, that cell shows a dash (—), not 0%. Where there is no goal-matched revenue, the value cell shows a dash, not $0.00. A dash means “no data to attribute here” — it never means “zero sales”.

A page that engines cited but that hasn’t earned a matched visit yet still appears in the table, with zeros and a low confidence band. When no cited page has a matched visit yet, you’ll see a short note — “Cited, but no on-site match yet” — so you know the gap is “still collecting”, not “nothing happened”.

Site-wide Audience

Below most-cited pages, the Audience panel breaks down who arrives from AI answers, pooled across every tracked prompt. Instead of a wall of small tables, it is one panel with three question-first views you switch between:

• Location — which countries (and, where your privacy mode allows, cities) your AI-referred visitors come from.
• Platform — device and browser.
• Source — which answer engine or referrer sent them, and any campaign tags.

You can drill into a row in place — pick a country to see its cities, then step back with the All countries control — without leaving the panel.

Two honest behaviours to expect:

• Below the floor. The breakdown only appears once enough AI visits have been attributed to be meaningful. Under that floor you’ll see a single sentence — “Audience breakdown appears once about N visits are attributed to AI answers. M so far this period.” — rather than a misleading sliver.
• Privacy-mode limits. Finer detail (city, browser version, full referrer URLs) needs the detail your site’s privacy mode collects. On Mode A sites those drills show a Mode B/C note instead — this is expected, not an error. See Privacy modes explained.

If a single view is too sparse to break down for the period, it says so rather than showing noise; the other views stay put.

The per-prompt answer view

Choose View analysis on any prompt (from the outcomes table) to open its answer view at /citations/[id] — the full story for one prompt. It is built to be honest about what each engine actually did.

What the engines answered

For each engine, over the period, you see its outcomes as distributions, never a single rank: how often you were cited, mentioned, or absent; where you tended to sit (Top 3 / 4–10 / below 10); and the sentiment spread. When you weren’t cited, a plain sentence names the honest reason — and this is where the three-state retrieval honesty matters:

• In the retrieval set, not cited. The engine did consider your domain and chose other sources. This is the actionable state — content and authority work can move it.
• Never surfaced. Your domain didn’t appear in the engine’s exposed retrieval set for these prompts at all.
• Retrieval set not exposed. Some engines don’t reveal what they considered. We won’t claim “never surfaced” without evidence, so we say plainly that we can’t see it.

Each “not cited” state links to the sources that were cited instead, so you can see who took the answer.

Sources cited for this prompt

A ranked list of every domain the engines cited in their answers for this prompt, ordered by how often and how high they were cited. Your own row carries a You badge. This is the surface where you decide a source matters enough to track as a competitor — see Tagging competitors.

Cited pages, position, and the observation timeline

• What these citations drove — the same cited-pages join as the Analytics page, but scoped to this one prompt.
• Cited position — shown as a distribution (how often you appeared at position 1, 2, 3-and-beyond), never a single “AI rank” that wouldn’t survive the next check.
• Sample answers — one sample per engine, where the provider’s terms let us show it, pattern-redacted (emails, phones, IPs stripped). For an engine we can’t quote you’ll see an “Excerpt unavailable for this engine” note. On days an engine rolled out a fleet-wide model change, a Platform model update marker reminds you that swings reflect the platform, not your performance.

Tracked prompts

Open the Tracked prompts page to manage what we watch. A tracked prompt is a question you want to watch across the AI answer engines — for example best privacy-first analytics tools or how do I install an analytics script without cookies. We poll the engines for each one on a schedule and record whether they cite you, mention you, or leave you out.

The page has three parts: your tracked prompts list, any suggested prompts we auto-discovered, and the Competitors tagging list.

Managing prompts

Choose Add prompt to open the dialog. You set three things:

• Prompt — the exact wording to watch.
• Engines — which answer engines to check (all your plan’s engines selected by default).
• Check frequency — how often we poll (see cadence below).

Choose Start tracking and we begin watching. If you try to add a prompt you’re already tracking, we tell you instead of silently doing nothing. A prompt you tracked before and removed cannot be re-added.

On each existing prompt you can Pause it, Resume it, Edit settings (engines and cadence — the question text stays fixed once tracked), or Remove it. Removing frees up a slot against your organization’s shared allowance. Suggested prompts are ones we auto-discovered; choose Accept to start tracking, or Dismiss. Suggestions don’t count against your allowance until you accept them.

You can only manage prompts if you are a member or above — viewers see the lists but not the controls.

Check frequency — Auto-pilot, Weekly, or Daily

Each prompt runs in one of three check modes:

• Auto-pilot (the default) — we choose how often to check it for you, stepping the frequency up or down as the prompt’s results settle or move. Leave it here unless you have a reason to pin it.
• Weekly — pin the prompt to a weekly check.
• Daily — pin the prompt to a daily check. Daily is a paid feature (Pro and above), and each plan allows only a set number of Daily prompts at once — its Daily slot cap.

Two different limits can stop you choosing Daily, and the dialog tells you which one:

• Daily isn’t on your plan. On Free, Starter, and Growth, Daily is locked with the note “Daily checks are available on higher plans.” Upgrade to Pro to unlock it.
• You’re at your Daily limit. On a Daily-capable plan, once your Daily slots are full the dialog says “You’re at your daily limit — switch one to Weekly or upgrade.” Switch an existing Daily prompt back to Weekly to free a slot, or upgrade for more.

Your Daily slots — like your prompt allowance — are shared across every site in your organization (see Allowances are per organization).

Tagging competitors

The Competitors list on the Tracked prompts page is where the word “competitor” finally earns its place. Every domain an engine cites beside you starts as a plain cited source — we make no assumption about whether it’s a rival. When you decide a source actually competes with you, you tag it:

• On the answer view’s Sources cited for this prompt list, choose Tag as competitor, or add one directly on the Competitors list with Add competitor.
• Paste any URL — we keep just the registrable domain (https://www.plausible.io/pricing → plausible.io). You give it a display name (e.g. “Plausible”) — that name is how it reads on your surfaces.

Once tagged, that source:

• Wears a Competitor badge wherever it appears (the outcomes table’s “cited instead” column, the answer view, the cited-sources list).
• Powers the Competitors lens on the Share of voice panel and the gap ranking.

Tagging changes nothing about polling — it’s purely a label on data you already have. One tag per domain; removing a tag reverts that domain to a plain cited source everywhere, instantly. A tag you add before the domain has ever been cited sits dormant and starts matching the moment it first appears in an answer.

Your competitor labels are private to your own dashboard — see the note under Methodology.

Allowances are per organization

Two things are worth knowing about how your limits are counted:

• Your allowance is one shared pool, not per site. The number of tracked prompts and the number of Daily slots your plan includes are counted across all the sites in your organization at once — not separately per site. If your plan includes 120 prompts and 30 Daily slots, that’s 120 prompts and 30 Daily slots to spread however you like across your sites, not 120 per site.
• Each plan also caps how many sites you can have (Free 1 · Starter 3 · Growth 10 · Pro 25 · Scale 50). At the cap, adding a site prompts an upgrade (or, on the top plan, a note to contact us).

When you reach your prompt allowance, the Add prompt button is replaced by an Upgrade plan link (if you can change the plan), and the usage bar notes “You’ve used all N tracked prompts on your plan.” Remove a prompt you no longer care about — on any of your sites — to free a slot, or upgrade for more.

If you change plans

Your prompts are never deleted when you change plans — your settings are remembered:

• Downgrade. If your new plan allows fewer Daily prompts than you’re using, the extra ones keep their place but are checked Weekly for now. We email you, and each affected prompt shows a “Daily — resumes when you upgrade” note in the list. Nothing is lost; the moment you upgrade again, they return to Daily automatically. (You can also choose which prompts stay Daily if the default isn’t what you want.) Engines that aren’t on your new plan are paused the same way — and switched back on when you upgrade, exactly as you had them.
• If a downgrade leaves your whole organization over its shared pool, citation monitoring for the organization pauses (read-only — your existing data still shows) until you upgrade or retire enough prompts to fit. We tell you when this happens.
• Upgrade. Everything that was suppressed for plan reasons is restored to exactly how you had it — Daily prompts go back to Daily, plan-paused engines switch back on. Engines you turned off yourself stay off.

Methodology

This section explains the rules behind every number.

Visit grain, not unique people

The conversion and revenue figures attributed to a citation are counted at visit grain — one matched visit (a page, an AI source, and a time window), not one unique person.

On Mode A sites there are no visitor or session IDs, so we cannot tell two visits by the same person apart. The figure is correct as a count of attributed visits, and it is the right signal for which citations drive outcomes — but read it as a guide, not as a precise per-person rate. It stays correct at visit grain on Mode B too; session-accurate de-duplication arrives when Mode B’s session IDs ship. We label the figure visit-grain in both modes so the grain is never ambiguous. See Privacy modes explained for the modes.

We never invent a number to fill a gap. If a citation drove no goal-matched revenue, the revenue reads as a dash (—), not a zero — “no data to attribute here”, not “zero sales”.

Confidence, shown as a band

Attributing a visit to a citation is probabilistic — we match AI-referred visits to the page an engine cited, and the strength of that match varies. So we never show attribution as a bare exact number. Every attributed page carries a confidence band:

A citation we found but could not corroborate with a single matched visit shows a low band — honestly low, never hidden and never inflated to look certain.

Directional vs measurement

Each visibility figure wears an honesty label so you know how much weight it can bear:

• The “API-grounded / directional” chip — every result is grounded in a real API poll, and read as directional.
• A sample count — how many observations the figure rests on.
• A confidence / variance band — how wide the uncertainty is.

The credibility tier is chosen from that band, never from your plan:

• When the band is tight enough (at or inside our measurement threshold), the figure is a Measurement — stated plainly, with no caution badge.
• When the band is wider, the same figure is a Directional sample — it gets a caution badge, and the word “measurement” is deliberately absent.

The rule is the band, not the plan name: a well-sampled prompt on any tier earns “Measurement”; a thinly-sampled one is flagged as directional even if everything else is identical.

Position is a distribution, never a rank

When an engine cites you, where in its answer it places you varies from one observation to the next. So we never show a single “AI rank” number — a figure that looks precise but does not repeat. Instead, cited position is shown as a distribution: how often you appeared at position 1, 2, 3-and-beyond across the window. The takeaway is the shape, not a single digit.

”Source” vs “competitor”

We are deliberate about this word. A domain an engine cites beside you is a cited source — nothing more. It becomes a competitor only when you tag it as one (see Tagging competitors). So the default Share-of-voice lens and every “other sources are cited” message stay source-neutral; “competitor” appears only where you’ve made the call. Your competitor labels are stored per site and shown only on your own dashboard — never to other customers, never on a public dashboard, never in a CSV export.

Which AI engines we monitor

AI Citations watches a fixed set of AI answer engines — the ones we can poll directly for what they cite. These five:

Copilot and You.com can still send you visitors — you will see those on the AI Visibility tab — but we do not poll them for citations, so they do not appear here.

Empty and “collecting” states

This feature never shows a blank screen. Instead it tells you which stage you’re at:

• No tracked prompts yet. The verdict reads “No AI citations yet” and invites you to add a prompt. Nothing to watch means nothing to show.
• Collecting your first results. You’ve added prompts but the engines haven’t been polled yet. The verdict reads “Collecting your first results” — first results land within about a day. Check back, or widen the period.
• Cited, but no on-site match yet. Engines are citing you, but no visit has been matched to a citation. The pages still appear with zeros and a low confidence band, and a note explains the gap.
• Polling paused platform-wide. Occasionally you may see “Polling is paused platform-wide. You’re seeing the last data we collected.” This is a maintenance pause on our side, not a problem with your site — your most recent data still shows, and collection resumes automatically.

If a panel looks emptier than you expect, the most common cause is the period — try widening it.

Troubleshooting

“Daily” is greyed out when I add a prompt. That’s a plan limit, not a bug. Daily checks are a paid feature, available from Pro up (Pro and Scale). On Free, Starter, and Growth, prompts are checked Weekly or on Auto-pilot. Upgrade to Pro to unlock Daily.

“You’re at your daily limit.” Different from the above — your plan does include Daily, but you’ve filled all its Daily slots (counted across all your organization’s sites). Switch a prompt back to Weekly to free a slot, or upgrade for more Daily slots.

The conv. rate or est. value column shows a dash. That site has no goal data to compute the figure from. Conversion and revenue need goals configured in site settings. A dash means “no data to attribute”, never “zero”.

Engines are citing me but the visits column is zero. A citation and a matched on-site visit are two different things. The page is cited, but no visitor we could attribute to that citation has landed yet — the confidence band stays low until one does. This is normal early on.

The verdict says “Other sources are cited — you aren’t yet”. Who are they? They are whatever domains the engines happened to cite for those prompts — plain cited sources, not “competitors” unless you’ve tagged them. Open the prompt’s answer view to see exactly who was cited in your place, and tag any that genuinely compete with you.

The Share-of-voice “Competitors” lens is empty. That lens only shows sources you’ve tagged as competitors. If you haven’t tagged any yet, it’s empty by design — switch back to All sources, or tag a source on the Tracked prompts page.

The Audience panel says “Available in Mode B/C” (or the breakdown is missing). Either that view needs detail your site’s privacy mode doesn’t collect (Mode A sites can’t break visits down that finely), or too few AI visits have been attributed to break down yet. See Privacy modes explained.

An engine I expected is missing. We only poll the five answer engines listed above. Copilot and You.com can send you visitors but aren’t polled for citations — find their referrals on the AI Visibility tab.

Why isn’t the number exact? By design. Citation attribution is probabilistic, so we show confidence bands and directional labels rather than false precision. See Methodology.

Related

• The AI Visibility tab — AI agents crawling your site, and real people arriving from AI.
• The AI traffic channel — the “AI” channel in the Referrers report.
• Understanding AI agents vs visitors from AI — two different things, kept separate.
• Privacy modes explained — why counts here are at visit grain, and what each mode can break down.
• Understanding your dashboard — the rest of the dashboard.