Numbers beat adjectives: the 33% that AI citation research is built on

Updated: 2026-08-27

Quick answer: adding concrete numerical data to a page raises its visibility in AI answers by 33% to 41%, according to GEO-bench measurements from Princeton. Not "significantly improved results" but "conversion rose 23%". This is one of the few findings in AI search optimisation that comes from controlled measurement rather than from someone's opinion, and it is also the cheapest change on the list.

Why does a model prefer a number?

A retrieval system has to choose between passages that all sound plausible. It has no way to verify any of them. What it has is signal density.

A sentence like "our approach delivers substantial improvements" contains one claim and zero verifiable facts. A sentence like "response time dropped from 340ms to 90ms after we moved the cache layer" contains a claim, two measurements, a mechanism and an implicit method. When a model builds an answer, the second sentence gives it something to stand on and something to attribute.

There is a second, less obvious reason. Numbers make a passage self-contained. "Substantial improvements" needs the surrounding paragraphs to mean anything. "340ms to 90ms" survives being cut out of context, which matters because retrieval cuts everything out of context.

A model is built on numbers, and a number cannot lie. Eleven is always eleven, in any language and any context. An adjective needs someone to interpret it. That is why our pricing page scores higher than any other: there you simply cannot write in adjectives.

Information gain

What a page adds to what search already knows, as opposed to restating it.

GEO-bench

A measurement set used to compare techniques for preparing pages for generative search.

What does the measurement actually say?

The finding comes from GEO-bench, a benchmark built at Princeton to test which content changes move AI visibility. Researchers took the same pages, applied single modifications, and measured the change in how often generative engines cited them.

Adding statistics: 33% to 41% increase in visibility over baseline.

A related finding from the same body of work: readers reported 37% higher subjective trust in text containing concrete figures. That second number matters commercially even if AI never sees the page.

Two things are worth noting about how this is usually reported.

It is a relative increase, not an absolute guarantee. A page that nobody cites will not start getting cited because you added percentages. The lift applies to pages already in the retrieval pool.

It was measured on original statistics. Copying three industry figures from a competitor's blog is not what was tested. The mechanism relies on the page carrying information that is not already everywhere.

Which rewrite takes ten minutes?

Take any page you have written and search it for these words: significantly, substantially, dramatically, considerably, greatly, highly, extremely, various, numerous, several.

Each one is a place where you had a number and threw it away.

Before: "We significantly reduced page load time, which greatly improved the user experience and led to a substantial increase in conversions."

After: "Page load dropped from 4.5 to 1.8 seconds. Bounce rate on mobile fell by a third, and checkout completions rose 12% over the following month."

The second version is the same length. It contains four numbers, a time frame and a segment. It can be quoted. It can be checked. It can be argued with, which is precisely why it carries weight.

Where do you find numbers when you think you have none?

Most people believe their business has no interesting data. Almost always they have plenty and have not thought of it as data.

Time. How long does the work take? How long has the company existed? How long until a client sees a result? "We have been doing this for eleven years" is a number.

Volume. How many projects, clients, cities, languages, integrations? A number that is small can still be specific: "we work with six languages" beats "we work with many languages".

Cost and range. Price ranges, budget brackets, what a typical project costs. Many businesses hide this and lose both the trust signal and the citation.

Process. How many steps, how many revisions, how many checks. "Every article passes eleven checks before publication" is more useful than "we maintain high quality standards".

Failure. The most credible numbers are the unflattering ones. "Our own site scored 96 out of 100 and got zero citations" is a sentence people remember, and we have used it more than any success figure we own.

What is the trap on the other side?

Here is where this advice gets misused, and where we had to build a countermeasure into our own auditing engine.

If numbers raise citation rates, the obvious move is to stuff a page with them. Sprinkle percentages everywhere, add three expert quotes, pile on trust badges, mark up everything with schema.

This backfires, and there is measurement on that too. Research on brand bias in model recommendations found that overloading a page with trust triggers drops a challenger brand's breakthrough rate from 55% to 21.2% in Claude. The model reads the pile-up as manipulation, which is a reasonable inference: pages that try that hard usually have something to compensate for.

Our engine now carries an explicit flag for this. If a page shows five or more persuasion markers while the body text is thin and has no original images, the score takes a penalty. The check exists because the alternative is an optimisation guide that teaches people to build spam.

The rule that survives: numbers work when they are load-bearing. A figure that supports a claim you are actually making helps. A figure dropped in because an article told you numbers help does not.

What does this look like in a scoring model?

In our engine the signal is measured as density rather than presence: how many numerical facts appear per thousand words, counting percentages, sums, durations and counts.

Presence alone would be trivially gameable. One number in a two-thousand-word page would pass. Density means the page has to actually be built on facts.

The threshold sits at roughly five numerical facts per thousand words for full credit, scaling down from there. That figure is ours, derived from what well-cited pages in our sample looked like, and I will flag it honestly: it is a project decision informed by the research, not a number lifted from the research.

Of the 75 variables in our current model, only ten carry thresholds taken directly from published measurements. The statistics signal is one of them.

What does the spread between our own pages show?

We scanned five pages of our own site on 27 August 2026 with the v1.2 engine. The technical half of the score barely moves. The content half does all the work.

Thirteen points separate the best content score from the worst. Three points separate the best technical score from the worst. Same site, same templates, same markup, same infrastructure.

The pricing page scores highest on content for an unremarkable reason: it is the page where we were forced to be specific. Prices are numbers. Tiers are counts. What each package includes is a list of concrete items. We did not optimise that page for AI. We just could not write it in adjectives, because nobody buys "affordable, flexible pricing" without asking what it costs.

The blog index scores lowest for the mirror reason: it is a list of links with promotional framing around it. Nothing on it is measurable.

The lesson is not "add numbers to your blog index". It is that the pages where your business is forced into specifics already outperform the pages where it is not, and that gap exists before anyone touches AI optimisation.

Frequently asked questions

Do I need original research to benefit? No, but you need original framing. Citing an industry figure with a link is fine. Citing five figures everyone else already cites adds nothing your competitors do not have.

What about pages where numbers feel wrong, like an About page? Those pages have the best numbers and use them least. Founding year, team size, cities covered, projects delivered. An About page without a single figure reads as a brochure.

Does this help with Google as well, or only with AI? Both, for different reasons. Classical search rewards content depth and dwell time; concrete figures help there indirectly. AI retrieval rewards fact density directly.

How precise do numbers need to be? Precise enough to be checkable. "About 200 clients" is weaker than "214 clients since 2015" and stronger than "many clients". Ranges are fine when the honest answer is a range.

Should I update numbers as they change? Yes, and date them. A figure with a date attached survives ageing gracefully. An undated figure quietly becomes a liability.

Is there a point where a page has too many numbers? Yes, when they stop supporting claims and start replacing them. A paragraph that is four statistics and no argument reads like a data dump and gets treated as one.

The short version

Adjectives are what you write when you have not measured anything. That was always true, and it was always a weakness in persuasive writing. What changed is that a system now reads your page and decides in milliseconds whether it contains anything worth quoting, and adjectives give it nothing to work with.

The fix is not stylistic. Go find the numbers you already have and put them where the claims are.

Sources: GEO-bench measurements (Princeton University); research on brand bias in LLM recommendation systems; project research digest compiled from 165 sources on AI search, August 2026. Details in sources/research-notes.md.