A number survives extraction when the sentence that carries it still says what was measured, in what unit, when, and according to whom, after everything around it is gone. “It dropped 40% last year” fails four ways once it leaves the page: dropped what, 40% of what base, which year, says who. “Median page weight on the sample fell from 2.4 MB to 1.4 MB between March 2025 and March 2026, according to our crawl of 1,000 retail sites” survives all four. (Both sentences are illustrative.)
This matters more for figures than for any other kind of text. A number is the part of a passage most likely to be quoted, and also the part that depends most on context the extractor did not keep. An answer engine lifts a passage, not a page, so the unit in a table header three screens up, or the date at the top of the post, is not part of what gets reused.
This post covers the rules for writing figures, units, ranges, and dates so they stay correct out of context, how to source a statistic in the same sentence, and how to find the vague ones across a site. It sits under Answer Engine Optimization: a page-level guide, and it is a close sibling of What makes a page citable, which covers why statistics and sources matter in the first place.
Why do numbers break when they’re lifted?
Because most of what makes a number meaningful lives outside the sentence. Writers put the unit in a column header, the date in the byline, the base in the previous paragraph, and the source in a footnote. A reader on the page reassembles all of that without noticing. A passage pulled into an answer has none of it.
Chunking: why passages have to stand alone makes the general case for self-contained passages. Numbers are the strictest version of it, because an unresolved reference in ordinary prose makes a sentence vague, while an unresolved reference in a figure makes it wrong. “Up 12%” quoted without its period can be read as current when it describes a year that has passed.
The GEO paper (Aggarwal et al., KDD 2024) tested adding statistics and citations to source text and reported that its methods could boost visibility in generative engine responses by up to 40%. That was measured on the authors’ own benchmark and engines, and it says nothing about whether a figure written badly helps. The point here is narrower: if figures are part of what gets lifted, they should be written to survive it.
What does a self-contained figure need?
Five parts, and most vague figures are missing two or three:
| Part | Question it answers | Weak | Self-contained |
|---|---|---|---|
| Subject | What was measured? | “it rose" | "average order value rose” |
| Value and unit | How much, in what? | “by 12" | "by 12 %” or “by USD 14” |
| Base or comparison | Compared to what? | “12% higher" | "12 % higher than Q2 2026” |
| Time | When? | “recently" | "in Q3 2026” |
| Source | According to whom? | (none) | “according to the company’s Q3 report” |
Not every sentence needs all five. A figure that is plainly a fixed property, such as “the controller weighs 3.2 kg,” needs subject, value, and unit; time and source are implied by the product page. But any figure that changes over time, compares two things, or comes from someone else needs all of them, and needs them in the same sentence.
How should units be written?
Write the unit every time, next to the number, in the same form throughout the page. NIST’s style guide for SI units sets out conventions worth adopting even outside scientific writing:
- A space between the value and the unit symbol:
3.2 kg,25 °C. The guide’s exception is plane-angle symbols such as degrees of arc. - A space before the percent sign in the NIST convention:
0.25 %. Many style guides drop the space; pick one form and keep it. - Make it clear which unit each value belongs to. NIST’s example is
0 °C to 100 °C, and it recommends the word “to” for ranges instead of a dash. - Keep information out of the unit. NIST accepts “the Pb content is 5 ng/L” and rejects “5 ng Pb/L.” Put what was measured in the words, not in the symbol.
For extraction, the most important of these is the third. A table cell that says 12–15 with the unit in the header becomes two bare numbers once the header is dropped. A cell that says 12 kg to 15 kg does not depend on the header. The same applies to currency: write USD 49 or $49 per seat per month, not 49 under a “Price” column.
Units that hide inside abbreviations deserve a full form on first use. “TB” in a storage spec, “kW” on an energy page, and “MAU” on a SaaS page are clear to an expert and opaque to a general answer. Define them once in a sentence that also carries the number.
How should dates be written?
Use absolute dates in the visible text, and machine-readable dates in the markup. Relative time is the most common way a correct sentence becomes wrong after extraction.
| Avoid | Write instead |
|---|---|
| recently | in September 2026 |
| last year | in 2025 |
| this quarter | in Q4 2026 |
| currently | as of 11 October 2026 |
| over the past few years | between 2022 and 2026 |
| new | released in March 2026 |
“As of” is the pattern to reach for with any figure that changes: prices, counts, rankings, support policies. It costs three words and turns a statement that will silently expire into one that is true permanently.
For the visible date format, pick an unambiguous one. 11 October 2026 and 2026-10-11 read the same everywhere; 10/11/2026 means different days in different countries. In the markup, use the HTML <time> element with a datetime attribute, which MDN documents as translating dates into machine-readable form, using formats such as YYYY-MM-DD:
<p>Prices on this page are <time datetime="2026-10-11">as of 11 October 2026</time>.</p>
In structured data, Google’s Article documentation specifies datePublished and dateModified in ISO 8601 format and recommends including time zone information. Freshness signals and honest dateModified covers when those dates should change; the rule here is only that the dates inside the text should be as explicit as the ones in the markup.
How should ranges, approximations, and comparisons be written?
State both ends of a range with units, mark approximations with a word instead of false precision, and name both sides of a comparison.
- Ranges: “between 40 and 60 words,” “from 2 kg to 5 kg.” Not “40–60” in a sentence where the unit is two clauses away.
- Approximations: “about 3,000 requests per hour” is honest. “2,987 requests per hour” from a one-day sample implies precision the measurement does not have.
- Comparisons: “twice as fast as the previous model” needs the previous model named. “Faster than competitors” cannot be checked or quoted.
- Percentage versus percentage points: a rate that moves from 10 % to 12 % rose by 2 percentage points, or by 20 %. Say which.
Tables help, as long as the cells carry their own units or the headers say them plainly and the table sits under a heading that names what it measures. Tables and lists that survive extraction covers table structure.
How do you quote a statistic someone else published?
Put the source in the same sentence as the number, link it, and state the scope the source measured. A statistic with its source in a footnote arrives in an answer with no source.
A pattern that works:
According to [source], [subject] [value with unit] [time or sample], [what it was measured on].
The scope clause is the one most often dropped, and it is the one that keeps a figure honest when it travels. A survey of a few hundred marketers is not a measurement of the web. A study of one engine’s answers is not a measurement of all engines. Say what was measured and let the reader decide how far it generalizes.
Two things not to do: round a source’s figure into a stronger claim (“nearly half” for 41 %), and cite a secondary article that cites the study when the study itself is available. Link the primary source. The broader case for citing sources is in Citability and entity signals: a GEO guide.
How do you check one page by hand?
Read every number on the page and ask, for each one, whether the sentence it sits in answers the five questions in the table above. Then search the page text for relative time words: “recently,” “currently,” “this year,” “last year,” “new,” “now,” “today.” Each hit is either fine because an absolute date sits in the same sentence, or it is a fix.
Paste the passages that carry your most important figures into the answer block checker with their headings, to see whether each still answers its question when read alone.
How do you find vague figures across a site?
Use a crawler’s custom extraction or search feature with a pattern for relative time words, and export the matches to a spreadsheet. Most desktop crawlers can search page text for a regular expression and report a count per URL. A pattern like this one catches the common offenders:
\b(recently|currently|this year|last year|this quarter|last month|nowadays|these days)\b
Sort the export by count, then filter to your highest-value templates. A template that renders “Updated recently” on every page is one fix that clears thousands of hits. For units, a second pattern finds bare percentages and currency with nothing nearby, but it produces more false positives; treat it as a reading list, not a defect list.
If you have a local copy of the site’s built HTML and a shell, the same search runs as:
grep -rEoi '\b(recently|currently|this year|last year)\b' ./dist --include=*.html | cut -d: -f1 | sort | uniq -c | sort -rn
How GEOBee reads figures
GEOBee’s GEO lens scores Citability at 25 of 100 points, and statistics are one of the things it looks for, alongside outbound citations and quotable blockquotes. The three lenses lists the signal and its weight. It counts whether figures and sources are present; it does not judge whether a figure has its unit, its date, or its base. That part is the read described above.
To see which pages have no figures or sources at all, sorted by template, download GEOBee and run a quick scan during the 7-day trial.