What is AI citation velocity?

AI citation velocity is the rate of change of your AI citation rate over time: how fast the percentage of topic queries where AI engines cite your site is climbing or falling. It is measured by running the same fixed query panel against the same AI platforms on a repeating schedule and recording the delta between runs, expressed in percentage points per quarter. A site at a 5% citation rate climbing 5 points per quarter is on a stronger trajectory than a site flat at 15%.

How do you calculate AI citation velocity?

Velocity = (citation rate this period minus citation rate last period) divided by the interval between measurements. Measure citation rate the same way both times: identical query panel, identical platforms, cited responses divided by total responses. If you measured 5% in April and 15% in July, your velocity is +10 percentage points per quarter. The panel must stay identical between runs; change the queries and you have a new baseline, not a velocity.

What is a good AI citation velocity?

There are no published industry-wide velocity benchmarks yet, so calibrate against your own baseline. From our tracked domains: flat velocity at a near-zero rate signals a structural problem (fix infrastructure before content), any sustained positive velocity of 3 or more percentage points per quarter on a 20-query panel means the trajectory is working, and a negative swing of 5 or more points in one period usually traces to a model update rather than anything you changed. Verify against announced model releases before reacting.

AI citation velocity vs citation rate: which matters more?

For an established site, rate matters more: it measures how much AI answer share you hold today. For a newer or recently restructured site, velocity matters more: rate lags structural fixes by one or two crawl and index cycles, so a low rate with strong positive velocity is a leading indicator that the fixes are compounding. Track both, but weight velocity when your site is less than a year into AI visibility work.

Can AI citation velocity be negative?

Yes, and it has three common causes: an AI model update that shifted source preferences (can move rates 20 or more percentage points with no change on your site), a technical regression such as a robots.txt change, a move to client-side rendering, or schema removed during a rebuild, and competitive displacement, where a stronger page replaced yours in the engine's source selection. Diagnose in that order: check announced model releases against your measurement window first, then recent site changes, then who now holds your core answers. One negative quarter aligned with a model update is noise; two consecutive negative quarters with no model change is a real decline.

AI Citation Velocity: Formula, Benchmarks, and a 90-Day Example

Two sites both show a 5% AI citation rate today. One was at 5% a year ago and is stuck. The other was at 0% ninety days ago and is compounding. A snapshot metric cannot tell them apart, and that failure has a cost: the stuck site needs a structural rebuild while the climbing site just needs patience, and mixing up those two prescriptions wastes a quarter either way. The metric that separates them is AI citation velocity: the rate of change of your citation rate, measured against a frozen query panel. This post defines it precisely, gives the formula and a real 90-day example, and leaves you with a tracking method you can run quarterly in about an hour. It builds on the framework that treats getting cited by AI engines as a measurable property of your site, and on the 20-query protocol from How to Measure Your AI Citation Rate.

Why a Snapshot Metric Misleads New Sites#

Citation rate is the percentage of topic queries where an AI engine cites your domain as a source. It is the right north-star metric, but it lags. Structural fixes (schema, answer-first restructuring, crawler access) take one or two crawl and index cycles to show up in AI answers, and training-data recognition lags by months. A site that fixed everything last month can still measure 0% today.

Our own baseline made this concrete. When we audited seven domains while building the AVR framework, chudi.dev measured a 0% citation rate: no per-page structured data, no FAQPage or HowTo schema, content that answered questions but not in an AI-readable format. A snapshot said "invisible." What the snapshot could not say was whether the fixes then underway would compound. Only the second measurement, against the same query panel, could answer that.

The Formula#

Velocity is the delta between two same-panel measurements, divided by the interval:

citation_velocity = (rate_now - rate_previous) / interval
 
# April: 1 citation  / 20 queries = 5%
# July:  3 citations / 20 queries = 15%
# velocity = +10 percentage points per quarter

Three rules make the number trustworthy:

The panel is frozen. Same 20 queries every run. Change the panel and you have a new baseline, not a velocity.
Per-platform, always. ChatGPT Search, Perplexity, and Claude cite differently and move independently. A +10 on Perplexity can hide a -5 on ChatGPT in a blended number.
Percentage points, not percent change. Going from 5% to 15% is +10 points. Reporting it as "+200%" makes small movements at low rates sound like explosions and becomes meaningless at 0%.

Two Velocities: Rate and Count#

Rate velocity, above, is the primary metric because it is normalized by your panel size. There is a second, complementary read: citation count velocity, the change in raw citations your site earns per period on a platform that reports first-party numbers. Bing Webmaster Tools' AI Performance tab is currently the only such surface, and it covers Microsoft Copilot only.

The count metric is what made our own 90-day example legible. In the freeCodeCamp guide to measuring AI citation rate, we documented the same site that measured 0% at the April baseline reaching 671 verified Copilot citations over the following 90 days, pulled directly from that tab. The domain's authority barely moved in that window. The structure work compounded faster than the authority work could, and it was the velocity read, not the snapshot, that showed the fixes were working weeks before rate velocity confirmed it.

Use count velocity as the early-warning channel (it updates continuously and costs nothing) and rate velocity as the quarterly verdict.

What Good Looks Like#

Honest caveat first: no one has published industry-wide velocity benchmarks, and anyone quoting a universal "good velocity" number is guessing. What we can offer is the calibration we use on our own tracked domains, on a 20-query panel:

Reading	Interpretation
Flat velocity at a near-zero rate	Structural problem. Infrastructure is blocking citation; more content will not move it. Diagnose before publishing anything else.
+3 or more points per quarter, sustained	The trajectory is working. Keep the cadence; do not change the panel to make the number look better.
Negative 5 or more points in one window	Check announced model updates inside the window before changing anything. One aligned negative quarter is weather.
Two consecutive negative windows, no model change	Real decline. Re-run the infrastructure scan and check whether a competitor now owns your core answers.

At a 20-query panel size, remember the confidence interval: a measured 15% has a true range of roughly 5% to 36%. Velocity inherits that noise, which is why single-window swings are annotated, not acted on, and two-window trends are what earn a response.

When Velocity Beats Rate#

Weight velocity over rate in three situations. First, a site less than a year into AI visibility work: rate lags fixes, so velocity is the leading indicator. Second, immediately after a structural overhaul: velocity tells you within one window whether the rebuild is compounding, while rate takes two or three windows to reach respectable absolute numbers. Third, competitive tracking in a young category: markets where every player has a low rate are decided by who is compounding, not who is ahead this month.

For an established site with a mature rate, the weighting flips back: defend the rate, and treat velocity as the alarm wire.

What to Do Next#

Velocity needs a baseline, and the baseline needs clean infrastructure to be worth trending. Run the free scan to check the 10 structural signals that cap citation at zero, fix what it finds, then freeze your 20-query panel and take measurement one. Ninety days later you will have the first number a snapshot could never give you. For the category context that explains why answer engines reward this kind of measurement discipline, see Answer Engine Optimization Explained on chudi.dev.