A Citation Cannot Check the Arithmetic: How to Audit Numbers in AI Search Answers

An AI answer can cite the correct specification sheet and still calculate the wrong saving, capacity, or payback period. A source supports its published inputs; a derived conclusion also depends on which inputs were selected, how their units were interpreted, and which calculation was performed. GEO content that includes numerical comparisons should expose those steps in a form a buyer can inspect. This article separates factual support from computational support, works through a hypothetical equipment comparison, and explains how to publish calculations that remain useful when an assistant changes the wording, converts units, or combines figures from several sources.

The number may not exist in the cited document

A product document might publish power consumption. A tariff page might publish an electricity price. A buyer might supply operating hours. The answer then calculates an annual cost difference.

That final number is a new claim. None of the three inputs alone proves it.

The answer needs a defensible connection between the selected source values and the derived result. A citation attached to the final sentence does not automatically provide that connection. The source may contain the necessary input but use a different test condition, product revision, or denominator.

For businesses, this matters whenever a buying decision includes a calculation: estimated energy use, shipping quantity, installation capacity, cost per deployment, service hours, or a performance comparison. The brand may be accurately named while its commercial value is overstated.

The review should ask, “Can another person reproduce this number from the stated evidence?” That is a different question from whether the answer contains a clickable source.

Separate five stages of numerical reasoning

The following decomposition is a practical audit method.

A calculator can help with the fourth stage. It cannot independently resolve the other four.

This distinction is consistent with the approach in *PAL: Program-aided Language Models*, which separates language-model problem decomposition from execution in a runtime. *Program of Thoughts Prompting* similarly delegates computation to an external program. Their benchmark findings support studying that separation, not assuming every tool-assisted commercial answer is correct. PAL, Program of Thoughts.

If an assistant generates code using the wrong operating hours, executing the code faithfully produces a faithfully calculated wrong scenario. The audit still needs to check the inputs and the intended task.

A worked example: power reduction and annual savings

The following comparison is hypothetical. All figures were chosen to illustrate the calculation; they do not describe a real product or a Xindar client.

Assume two pieces of equipment perform the same task under the same stated operating conditions:

The absolute power difference is:

`4.8 kW − 4.1 kW = 0.7 kW`

The reduction relative to the existing equipment is:

`0.7 ÷ 4.8 × 100 = approximately 14.6%`

The annual energy difference is:

`0.7 kW × 3,200 hours = 2,240 kWh`

The annual electricity-cost saving is:

`2,240 kWh × USD 0.14/kWh = USD 313.60`

The simple payback period based only on those electricity savings is:

`USD 1,200 ÷ USD 313.60/year = approximately 3.83 years`

That calculation is internally consistent. It is still conditional on equivalent work, valid average-power values, the stated operating time, and the tariff. It does not account for financing, downtime, maintenance, installation costs, taxes, or a changing load profile.

A publishable sentence would identify that scope: “Under the illustrative operating assumptions above, the annual electricity saving is USD 313.60 and the simple energy-only payback is about 3.83 years.”

“The product pays for itself in under four years” discards most of the conditions that make the calculation meaningful.

Check the denominator before accepting a percentage

In the example, 14.6% describes a reduction relative to the existing equipment. Dividing by the candidate's 4.1 kW instead produces approximately 17.1%. That figure answers a different question: how much higher the existing equipment's power is relative to the candidate.

Both calculations can be arithmetically valid. They cannot be used interchangeably.

The same issue appears in GEO reporting. If mentions rise from 20 of 100 answers to 30 of 100, the change is ten percentage points and a 50% relative increase. Saying “visibility increased by 10%” leaves the intended comparison unclear.

Another common ambiguity is an average of ratios versus a ratio of totals. Two markets with very different sample sizes should not be silently given equal influence merely because their individual percentages occupy two spreadsheet rows.

Publish the numerator, denominator, period, and comparison basis. If the answer changes the basis, the change should be visible. A precise percentage with an unnamed denominator is difficult to audit and easy to reuse incorrectly.

Units are part of the evidence

Power and energy are different quantities. A kilowatt is a rate; a kilowatt-hour measures energy. Multiplying power by time can yield energy under the stated assumptions. Dividing a purchase price by a power difference cannot yield a payback period.

Write units into the formula so their cancellation can be inspected. In the example, kWh multiplied by dollars per kWh yields dollars. Dollars divided by dollars per year yields years.

For industrial content, also state whether a specification is instantaneous, average, nominal, maximum, or measured under a particular test. A “maximum 4.1 kW” value cannot simply replace an average-power observation in an annual-use estimate.

Conversions need the same care. If a page offers both metric and US customary values, identify the underlying value and round only after the calculation. Two rounded display values can disagree slightly without indicating a product change.

A good comparison table includes units in column headings and conditions in adjacent notes. Placing all qualifications in a distant footnote makes it easier for a copied row to lose its meaning.

A tool result is an execution record, not a validation certificate

When an answer uses a spreadsheet, interpreter, or calculator, the audit should retain the expression or code and the selected inputs where they are available.

The useful record is small: input values, their origins, formula, output, and assumptions. It should not require access to the assistant's private reasoning. Reproducibility concerns the calculation and supporting evidence, not a narrative claiming to reveal hidden thought processes.

The W3C PROV model offers concepts for describing entities, activities, and agents involved in producing information. Those concepts can inform a calculation record: source values are inputs, the calculation is an activity, and the output is a derived result. Using that vocabulary does not mean commercial answer engines will read your record or award it a ranking benefit. W3C: PROV-O.

For most companies, a visible worked example and a downloadable input table are more practical than implementing a complete provenance system. Start with the information a technical reviewer actually needs.

Show sensitivity without inventing a confidence interval

Suppose the illustrative average-power estimates are uncertain. The existing equipment may use between 4.6 and 5.0 kW, while the candidate may use between 3.9 and 4.3 kW.

Using the most conservative combination, the difference is 0.3 kW. Using the largest difference, it is 1.1 kW. Under the same time and tariff assumptions, annual savings range from USD 134.40 to USD 492.80. The corresponding simple payback ranges from approximately 8.93 years to 2.44 years.

This is a scenario range formed from chosen bounds. It is not a confidence interval. It does not assign probabilities to the endpoints or establish that every combination is physically achievable. If the two power values move together with workload, the extreme combinations may be inappropriate.

The example shows why a single payback number can hide a decisive uncertainty. A buyer may need a measurement at the intended workload before making a purchase.

For content, publish the assumption that most changes the decision. Readers benefit more from knowing that operating time is uncertain than from seeing an apparently precise figure carried to five decimal places.

Build a calculation page a buyer can reuse

A useful numerical page has three layers.

The first explains the question and the result in plain language. It identifies the relevant comparison and the conditions.

The second provides the inputs and formulas. A table shows the value, unit, source or assumption, date, and applicable model. A worked example demonstrates the calculation with clearly labeled illustrative values or approved real data.

The third explains limits and substitution. It tells a reader which inputs can be replaced, which assumptions must remain comparable, and when the result is unsuitable.

If a calculator is offered, the page should still explain what it calculates. A widget that outputs “annual savings” without identifying baseline, workload, or excluded costs creates a new ambiguity. Preserve the submitted inputs with an exported result so another person can reproduce it later.

Keep sales claims separate from illustrative outputs. A demonstration using selected assumptions is a teaching example; it does not establish the savings every buyer will receive.

Audit an AI answer before it becomes sales material

Use the following sequence for material numerical claims:

1. Extract the exact derived claim. Identify the number, unit, comparison, and implied decision.
2. Find each input. Check whether it comes from a current source, the user, or an unstated assumption.
3. Confirm compatibility. Ensure the inputs describe comparable models, periods, conditions, and currencies.
4. Recalculate independently. Use a simple calculator or reviewed spreadsheet. Retain the formula.
5. Test the interpretation. Check denominator, baseline, exclusions, and whether the result is a total or a rate.
6. Examine decisive uncertainty. Vary uncertain inputs within justified scenarios rather than manufacturing probabilities.
7. Approve the wording. Make the sentence no stronger than the supported calculation.

Record a failure at the stage where it occurs. Rewriting the final sentence will not fix an obsolete specification. Changing a formula will not fix an invented usage assumption.

Xindar's answer-engine content service describes converting product knowledge and evidence into verifiable English content. Where a project includes numerical claims, a reproducible calculation record is a sensible deliverable to specify. Its usefulness can be reviewed directly even before any change in AI citations is observed.

Frequently asked questions

Is a cited calculation more trustworthy than an uncited one?

A relevant citation can support the input, but the formula and interpretation still require review. Citation presence alone does not validate the derived result.

Should every numerical page include code?

1. Straightforward formulas, units, and a worked example are often enough. Code helps when the calculation is complex, provided the inputs and assumptions remain visible.

Can we report a range instead of one number?

Yes, if the range has a defined basis. Distinguish chosen scenarios, measured variation, and statistical uncertainty rather than labeling all of them a confidence interval.

Does publishing formulas guarantee more AI citations?

1. It makes the calculation inspectable. Citation selection and the correctness of a later answer still need separate observation.

Sources and calculation status

Sources were reviewed on October 9, 2026. All equipment values and business scenarios above are synthetic teaching examples. The arithmetic is reproducible from the displayed inputs. Research findings concern the methods and benchmarks tested in the papers, not guaranteed accuracy in current AI search products.

• Gao et al.: PAL: Program-aided Language Models — separating problem decomposition from runtime execution.
• Chen et al.: Program of Thoughts Prompting — program-based numerical reasoning.
• W3C: PROV-O — a vocabulary for information derivation and provenance.