How AI engines actually answer. AI engines are answering your buyers' questions right now. This is the logbook of what they actually say.
Instrument · window · sample, stated next to every reading.
Winning the citation is not winning the answer. A third of the time the engine uses your page and drops your name.
This is the naming gap: an AI answer lists a brand's page in its sources without naming the brand in the answer text.
Every pooled per-instrument reading falls between 33% and 45%. AthenaHQ's monthly slices run wider, from 35.94% in October 2025 to 54.62% in December 2025, so the band applies to pooled readings only.
Per-run measurement beats per-prompt dashboards here: a prompt can be "cited" on some runs and not others, so averaging at the prompt level hides exactly the flip this page is built to show. No remedy is sold on this page. The phenomenon is the story.
A screenshot is not a measurement. Half of tracked prompts changed their answer while we watched.
Batched sweeps mean a flip can reflect an engine model update between sweeps, not day-to-day churn. This is an any-flip measure across a run series, not per-ask odds. It is why this page carries a changelog instead of a single snapshot number that never moves.
There is no single "AI search" to optimise for. Three engines read three different internets.
Part of the gap is a roster artefact: engines differ in whether they return sources at all. Even accounting for that, agreement stays low.
Kevin Indig's consensus gap (Growth Memo, 2026-05-11, 3.7 million citations) found 91.07% of cited URLs appear on a single engine and 2.37% on all three. Our reading points the same direction on a different unit, prompt-day-domain combinations rather than cited URLs, and we say so rather than presenting it as the same metric.
The answer is assembled from surfaces you do not control. Your website is one voice in the mix.
google.com is already the #2 cited domain in our tracker (1,830 citation entries) and youtube.com is #3 (1,096), measured 2026-08-21.
Half of everything AI engines cite is a fragmented long tail: 5,850 domains, each cited 60 times or fewer. Fragmentation is the norm, not the exception.
The #1 cited domain in our own tracker is our own domain, intelligentresourcing.co, at 2,942 of 50,421 entries. A brand's own tracker asks brand-relevant prompts, so its own domain over-samples. That is measurement circularity and we name it rather than hide it.
A number arrives with its instrument, window and roster attached, or it is not a measurement.
Of 74 AI-visibility tools censused by citedindex.com on 2026-08-15, only 6 publish a checkable methodology. This chapter is why we publish the way those 6 do.
Our instrument attributes most unnamed citations to Google AI Overview (52% to 73% unnamed). AthenaHQ's full history attributes most to ChatGPT (65.84%). Same aggregate, opposite attribution.
Same instrument, same metric: Google AI Overview ghosting reads 2.92% on a one-month slice and 21.98% across seven months. 7.5x apart on window choice alone.
Both instruments changed engines mid-flight, dated on this page. Leaderboards move when rosters move underneath them.
"Cited" can mean the answer carried sources (71.40%) or that the brand appeared among them (35.78%). A 2x gap sits inside one word.
The #1 cited domain in our own tracker is our own domain (2,942 of 50,421 entries). Every brand-run tracker carries this bias; we state ours instead of hiding it.
Every exhibit above is reproducible from the methodology: two phases, 851 base prompts, up to six engines, sweeps dated.
What to demand from any AI-visibility vendor, us included: the instrument, the window and the roster, stated next to every number. The honest counter-case: some of this divergence is real engine behaviour changing over time, not instrument error. We cannot separate the two from this data, and we say so rather than picking the reading that flatters us.
58.4% vs 14.2% citation rate at retrieval position 1 versus position 10.
Retrieval outperforms content tactics by roughly 7.6x; only 3 of 54 content-tactic tests were significant.
Heading-query match: 41% vs ~30%.
Brand mentions correlate with AI visibility at 0.664 versus backlinks at 0.218; YouTube mentions correlate at roughly 0.737.
Collected in two contiguous phases: a third-party tracker (AthenaHQ) from October 2025 to April 2026, and IR's own tracker from April 2026 onward, with a four-day overlap. Every figure on this page keeps its own phase, window and roster. Rates never blend across phases.
Sweeps are manually triggered on a non-fixed cadence, stated plainly rather than described as continuous tracking. Shares are upper bounds within a tracked set, never market share. A replication across two independent instruments is stronger evidence than either alone.
We publish an llms.txt file on this domain and claim nothing for it: three independent studies found it has near-zero effect on citation. We chose Dataset schema markup instead.
Every snapshot appends a row below. Movement is only claimed once the roster has held still across the compared snapshots.
An AI answer that lists a brand's page in its sources without naming the brand in the answer text.
Whether a tracked prompt's brand-mention state changes across repeated runs. Measured as any-flip across a run series, not as odds on a single ask.
The degree to which different AI engines cite different sources for the same question. Term attributed to Kevin Indig's Growth Memo research.
A share calculated only within the set of prompts and engines a tracker follows. It is an upper bound within that set, never a claim about total market share.
A logged answer is any AI response recorded for a tracked prompt. A cited run is a logged answer whose sources include a tracked domain. Not every logged answer is a cited run.
These five terms carry every number on the page. The FAQ answers the same questions using the same figures.
The questions people actually ask, answered with the same numbers.
The naming gap is an AI answer that lists a brand's page in its sources without naming the brand in the answer text. Across three independently run B2B trackers, 35.99% of cited runs fall into this gap (1,841 of 5,115, April to August 2026).
64.01% of cited runs name the brand; 35.99% do not, measured across three B2B trackers from April to August 2026. A second instrument, AthenaHQ, measured 44.49% unnamed on an earlier window, October 2025 to April 2026, on one B2B account.
Rarely. Only 8.90% of prompt-day-domain combinations show two or more engines citing the same domain (3,463 of 38,916). Kevin Indig's independent research found a similar pattern on a different unit: 91.07% of cited URLs appear on a single engine.
Not very. 53.46% of tracked prompts with five or more runs flipped their brand-mention state at least once, May to August 2026. Batched sweeps mean a flip can reflect a model update rather than daily churn.
Half of citations come from a fragmented long tail of 5,850 rarely cited domains. Agency and vendor sites make up 29.46%, and Google-owned surfaces plus YouTube together account for roughly 6%, based on 50,421 normalised citation entries.
On the AthenaHQ instrument, 71.40% of logged answers carried sources, while only 35.78% actually contained the tracked brand's domain among them. Those are two different numbers hiding inside the single word "cited."
In two contiguous phases: a third-party tracker (AthenaHQ) from October 2025 to April 2026, then IR's own tracker from April 2026 onward, with a four-day overlap. Sweeps are manually triggered on a non-fixed cadence across up to six engines.
Ask for the instrument, the window and the roster. Then read the leaderboard.
