Verifying an AI Citation Costs Three Quarters of a Cent

Grading 7,365 AI answer-engine citations against the specific claims they anchor cost $56.29.

That is the cumulative spend published by Answer-Source Fidelity, the Machine Relations instrument that pulls a cited page, isolates the sentence the citation sits beside, and grades whether the page actually supports it. The instrument publishes its own receipt alongside its results: 21,370 grader calls, $56.286115 spent, 15,925 citation events entered into the denominator, 7,365 of them graded — 46.2% of everything it looked at.

Divide it out and the unit cost is $0.0076 per graded citation. Three quarters of a cent to answer, for one citation in one AI answer, the only question that matters about it: does the page say what the answer says it says.

That number reframes the entire conversation about AI answer accuracy. The category has spent two years explaining why citation-level verification is impractical. It is not expensive. It is roughly the price of a paperclip.

What $56.29 actually bought

What the instrument did	Count
Citation events entered into the denominator	15,925
Events graded against their anchored claim	7,365
Grader calls issued	21,370
Cumulative spend	$56.286115
Spend per graded citation	$0.0076
Grader calls per graded citation	2.90

The 2.90 calls per graded citation is the interesting line. This is not one model asked once. Every citation is graded under an always-dual adjudication policy — two independent graders, and a judge model when they disagree — against a fixed five-label scale (Supported, Amplified, Contradicted, Misattributed, Fabricated), with a separate operational label isolating a page that could not be fetched from a claim that was actually fabricated. The grading models are pinned and published: Gemini 2.5 Flash and GPT-4.1 mini as the two graders, Claude Sonnet as the judge.

Three model calls per citation, at current published inference prices, is why the total lands where it does. The work is small-context and highly parallel: a claim, a page, a label. It is the cheapest kind of frontier-model work there is.

Per nightly cycle, the bill runs between $0.77 and $2.42. Eighteen collection-time cycles, running from September 2 to September 20, 2026, cost $30.67 between them. A year of nightly citation grading at that cadence prices out around the cost of a single mid-tier software seat.

The barrier was never the money

If checking a citation costs three quarters of a cent, then the reason a vendor's AI visibility dashboard reports how often you were cited but not whether the citation held up is not budget. There are only a few candidates left, and they are all more uncomfortable than cost:

The result is unflattering. Among citations that can be pinned to a claim, roughly half land at the top of the fidelity scale. Machine Relations publishes 54.8% supported for ChatGPT and 50.1% for Perplexity, each with its resolvable coverage attached. A product built on the premise that a citation is a win has a hard time shipping a feature that grades half of them as something less.

It makes the denominator visible. A citation count is a clean, large, growing number. A verified-support rate forces you to publish how much of your own data you could not check, which brings the next problem with it.

Verification requires a position on what "supported" means. Counting citations requires no epistemology. Grading them requires a label set, an adjudication policy, a reference set, and a published disagreement rate — each of which is a claim a competitor can attack.

None of those are engineering problems. They are willingness problems, and the price tag no longer hides them.

What the money cannot buy

The honest half of this finding is that spend does not purchase coverage. You cannot grade a citation whose page you cannot retrieve, and retrievability turns out to be the binding constraint.

On the frozen legacy corpus, 3,033 of 10,066 citation events were resolvable — 30.1%. On the collection-time series, which snapshots each cited page minutes after the answer is produced rather than days later, 4,332 of 5,859 were resolvable — 73.9%, and 51.7% of those resolvable citations were graded Supported. Same graders, same prompt, same claim map, same label set. Snapshotting at collection time more than doubles what you can check, and moves the resulting support rate by about three points.

And resolvability varies enormously by which engine produced the answer. Inside the legacy corpus, at one moment, under one method:

Engine	Resolvable citations	Resolvable coverage
ChatGPT	1,255 of 1,630	77.0%
Perplexity	1,298 of 2,726	47.6%
Gemini	392 of 4,349	9.0%
Claude	88 of 1,361	6.5%

Gemini and Claude sit below the instrument's evidence floor and are marked collecting; the instrument withholds a support rate for both rather than publishing one built on a tenth of the evidence. That withholding is the discipline, not a gap in it.

The spread is the point. Support rates, where they can be measured at all, converge near half — 54.8%, 51.7%, 50.1%. Coverage spans 6.5% to 77.0%. What separates one answer engine from another is not how well its answers are supported. It is how much of its evidence anyone outside the company can check.

That has a direct consequence for buyers: a cross-engine "accuracy" comparison built without stating coverage per engine is mostly a measurement of retrievability, dressed as a measurement of truthfulness.

The instrument's own disclosed limits

A receipt is only useful next to the error bars. Machine Relations publishes both, which is the reason its numbers are quotable at all:

• Inter-model agreement between the two graders is Cohen's κ 0.647, with raw agreement of 72.2%.
• The external calibration certificate — a sealed human-annotation evaluation, and the gate that actually determines publication eligibility — scores 83 of 91 items correct, an accuracy of 91.2%, with a Wilson lower bound of 83.6%.
• Against the public AttrScore benchmark (AttrEval-GenSearch, dataset revision 467dcdd2), built from the attribution-evaluation work published in Automatic Evaluation of Attribution by Large Language Models and released as open code and data, binary claim-support agreement is 55 of 60, or 91.7%, with a Wilson lower bound of 81.9%.
• The disclosed bias: on the benchmark's extrapolatory label — cases where the cited source implies the claim without stating it — agreement drops to 6 of 18, or 33.3%. The instrument is strong on attributable and not-attributable, and weak exactly where a human would also hesitate. It publishes that number beside the rates rather than under them.

A verification claim without figures of this shape attached is a marketing claim — and the FTC has been enforcing against unsubstantiated AI claims since 2024. This is the bar the category should be held to, including when we are the ones making the claim.

Five questions worth asking a vendor

The economics above make these answerable, not aspirational:

1. Of the citations you report for me, how many did you fetch and grade against the claim they anchor? A percentage, not a yes.
2. What is your resolvable coverage, per engine? If it is not broken out per engine, it is an average hiding a 70-point range.
3. What label set and adjudication policy? One model asked once is a guess with a confidence interval nobody computed.
4. What is your agreement rate against an external human-annotated benchmark, and where is it weakest? Every instrument has a weak label. The ones worth trusting say which.
5. When coverage falls below your evidence floor, what do you publish? "Not yet" is the right answer. A number is the wrong one.

The cost of running that verification is now public and it is $0.0076 per citation. Any vendor declining to do it is making a choice about what it wants to know, and that choice is no longer defensible on price.

Why this matters now

Answer engines are consolidating how business software gets shortlisted. Google has moved AI Mode into the core search experience and documents how AI features select and surface sources; the same shift is underway across every major assistant. The citation is becoming the unit of brand exposure, replacing the ranked link.

Treating that unit as self-evidently true is a measurement failure with commercial consequences. Frameworks like the NIST AI Risk Management Framework put verifiable evidence at the centre of trustworthy AI systems, and inference pricing across Anthropic, OpenAI and Google Cloud has fallen far enough that verification at citation level is an operating-expense rounding error.

The instrument, its receipts, its cohort partitions and its limits are published in full in The Citation Support Gap, with the underlying supply-side measurement in the Machine Relations Index. Our own working notes on what a defensible AI visibility measurement requires are collected on the AuthorityTech blog.

Sources and method

All figures in this piece are read from the live Answer-Source Fidelity public aggregate at https://machinerelations.ai/measurement/answer-source-fidelity/data.json, retrieved 2026-09-20, generated 2026-09-20T11:34:36Z. Spend, call count, event counts and outcome counts are the instrument's own cumulative block; per-cycle spend is the cycle_series block across 18 cycles dated 2026-09-02 to 2026-09-20. Coverage and support figures are quoted per cohort, per the instrument's own framing rule, which reports each measurement-semantics cohort as its own track and never blends them: the legacy baseline uses a grade-time fetch basis on a frozen corpus of 10,066 events; the collection-time series uses a collection-time snapshot with bounded retry and deferral across 5,859 events. Calibration, certificate and benchmark figures are from the calibration and benchmark blocks, benchmarked against AttrScore (AttrEval-GenSearch) at dataset revision 467dcdd2cd31f9b5e8625491f3bdf7af90943a8d. Model pricing pages for Gemini 2.5 Flash, GPT-4.1 mini and Claude were each retrieved live on 2026-09-20 and are linked inline above. No figure on this page is derived from a third-party AI visibility study.